research rail
Results, papers, essays, methods and rebuttals in alignment, interpretability, evals and deception detection that the watch loop flagged in the last 30 days. A separate screen question from the containment and governance rails (added 2026-10-18 after the Researcher sat on Nick Cammarata's "Neural networks are already clean" for three days). One per day can enter the morning brief as the Research signal; a picked signal gets a reading with an experiment design, a book note and a post draft automatically.
154 in 7 days · 176 in 30 days · corroboration = other flagged items in the window that share the subject.
| Day | Signal | Source | Also seen | Why flagged |
|---|---|---|---|---|
| 2026-10-23 | LoReFT GRAFT Better at Preventing RL Grift | lw-ai-tag lesswrong.com/posts/B7LnjyJQnkgwKdtTx/loreft-graft-better-at-preventin | Proposes a novel, parameter-efficient method (LoReFT) to improve upon an existing alignment technique (GRAFT) for reward hacking mitigation. | |
| 2026-10-23 | Simulated Control Evaluations | lw-ai-tag lesswrong.com/posts/uXiqyoKwpfjWfQXk5/simulated-control-evaluations | 1 | Describes a new method (SCEs) for testing AI safety monitors under conditions closer to real deployment. |
| 2026-10-23 | OpenAI Trains Frontier Models Offline After September Hacks Breached 3rd-Party Systems | zetik-com zetik.com/news/article/story_id-p008-226334 | 1 | Reports on changes in training methodology (offline/air-gapped) due to safety concerns following breaches. |
| 2026-10-23 | Gradient Routing Generalizes to Unrouted Hack Types | lw-ai-tag lesswrong.com/posts/ie7bSp5ZDwYkCQxuv/gradient-routing-generalizes-to- | 1 | Investigates generalization of gradient routing to different types of reward hacks. |
| 2026-10-23 | An alignment assessment of the UK AISI cybertesting incident | anthropic-research anthropic.com/research/alignment-assessment-ukaisi | Investigates concerning unsanctioned actions of a frontier model in a cybertesting incident. | |
| 2026-10-23 | I measured an AI sandbox's audit export. It contained zero records — so I built the layer that keeps its own. | dev-to dev.to/techdevfun/i-measured-an-ai-sandboxs-audit-export-it-contained- | Describes a new technical layer for creating verifiable, tamper-proof audit logs for AI agent actions within sandboxes. | |
| 2026-10-23 | What Is Mechanistic Interpretability? AI Neurons Explained | pod-nate-bjones local podcast transcript: 2026-10-22-what-is-mechanistic-interpretability-ai-neurons-e | Reports on a finding that 99-100% of neurons in an improved vision model's final layer are interpretable. | |
| 2026-10-23 | METR’s Red Team Broke OpenAI’s Agent Monitoring Three Ways. Each One Attacks a Different Assumption. | forkast-news forkast.news/metrs-red-team-broke-openais-agent-monitoring-three-ways- | 1 | Reports on red-teaming an internal agent monitoring system for potential misalignments. |
| 2026-10-23 | MIT Researcher Traces AI Hallucinations to 1950s Randomness Injection, Warns of Creativity Tradeoffs | zetik-com zetik.com/news/article/story_id-p008-226096 | Suggests a mechanistic link between historical design choices (randomness injection) and a core failure mode (hallucinations). | |
| 2026-10-23 | Simulated Control Evals | arxiv-cs-ai-agents arxiv.org/abs/2610.31593v1 | 1 | Introduces Simulated Control Evaluations (SCEs) to broaden control evaluation settings by using an LLM to simulate tool responses. |
| 2026-10-23 | Questioning Instrumental Convergence in RLVR: Checkpoint Evidence from OLMo-3 | arxiv-cs-ai-agents arxiv.org/abs/2610.31624v1 | Empirical claim about model behavior across different training stages (SFT to RLVR) regarding instrumental convergence. | |
| 2026-10-23 | EpiSEM-Bench: Evaluating Safety and Evidence Recovery in LLM Agents under Missing Evidence | arxiv-cs-ai-agents arxiv.org/abs/2610.31650v1 | Introduces an executable benchmark (EpiSEM-Bench) specifically designed to evaluate safety and evidence recovery in LLM agents when critical facts are omitted. | |
| 2026-10-23 | Data-Void Exploits in Web-Searching LLMs | arxiv-cs-cr-agents arxiv.org/abs/2610.30365v1 | Presents a novel vulnerability (data-void exploit) specific to how LLMs synthesize answers from web search, which is a concern about model control/reliability. | |
| 2026-10-23 | Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems | arxiv-cs-cr-agents arxiv.org/abs/2610.30418v1 | 1 | Empirical quantification of 'mosaic effect' leakage in LLM agents over confidential data. |
| 2026-10-23 | Beyond End-to-End Recovery: Towards a Better Understanding of LLM Steganography via Rolewise Decomposition | arxiv-cs-cr-agents arxiv.org/abs/2610.30516v1 | Presents a new framework (rolewise decomposition) to analyze LLM steganography by separating encoding and decoding failures. | |
| 2026-10-23 | Model-intrinsic, Weight-level, and Training-free Spectral Fingerprinting for Large Language Models | arxiv-cs-cr-agents arxiv.org/abs/2610.30626v1 | Proposes a novel, weight-level, and training-free method for model fingerprinting based on spectral structures of LLM parameters. | |
| 2026-10-23 | Amplifying Latency and Cost of LLM Inference via Rowhammer Attack | arxiv-cs-cr-agents arxiv.org/abs/2610.30798v1 | Presents a novel hardware attack (Rowhammer) targeting the compilation pipeline (Triton/LLVM) used for efficient LLM inference, demonstrating a method to degrad | |
| 2026-10-23 | AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.31366v1 | Presents a new benchmark (AfterDeleteBench) and empirical results demonstrating information persistence/recovery from deleted memory in LLM agents. | |
| 2026-10-23 | RSI Control: Can AI Models Perform Covert Sabotage in the Self-Improvement Process? | arxiv-query arxiv.org/abs/2610.27735v2 | 1 | Presents a new testbed (RSI Control Arena) and empirical results on covert sabotage during self-improvement. |
| 2026-10-23 | Model Organisms of Collusion: Corrective Transfer in Untrusted Monitors | arxiv-query arxiv.org/abs/2610.30582v1 | Empirical result showing that fine-tuning a model monitor can significantly reduce unseen collusion risks. | |
| 2026-10-23 | AgentPhishing: A Benchmark of LLM Email Agents Under Adaptive Attack | arxiv-query arxiv.org/abs/2610.31404v1 | Introduces a new benchmark (AgentPhishing) and method (PhishOpt) for evaluating LLM email agents against adaptive, persuasive phishing attacks. | |
| 2026-10-23 | Behind Aggregate Scores: What Systematically Varying Arena-Hard's LLM Judge Reveals | arxiv-cy arxiv.org/abs/2610.30512 | Empirical claim about how LLM judges' verdicts are unstable based on minor changes in instructions/setup. | |
| 2026-10-23 | Probing and Steering for Safety-Relevant Behaviours in Large Language Models | arxiv-cy arxiv.org/abs/2610.30363 | Presents a systematic empirical study on probing and steering multiple safety-relevant behaviors using activation probes in detail. | |
| 2026-10-23 | Detecting AI Training in Datacentre Power | arxiv-cy arxiv.org/abs/2610.30182 | Presents a novel, technical method for detecting large-scale model training activity using physical infrastructure monitoring (power consumption). | |
| 2026-10-22 | MIT’s 13 Eko Agents Invent Games and Outperform Rivals After 30 Hours of Play | zetik-com zetik.com/news/article/story_id-p008-226083 | 2 | Empirical demonstration of emergent, complex behavior from unguided agent interaction. |
| 2026-10-22 | Left alone with nothing to do, AI agents started playing | techcrunch-ai techcrunch.com/2026/10/22/left-alone-with-nothing-to-do-ai-agents-star | 2 | Describes emergent complex behavior (inventing games) from unconstrained AI agents. |
| 2026-10-22 | Suddenly, Unclean | lw-ai-tag lesswrong.com/posts/8ozpJjwi5wjiyvuvd/suddenly-unclean | Proposes a theoretical discontinuity in the interpretability/cleanliness of neural networks based on model scaling. | |
| 2026-10-22 | LLMs are making formal software verification feasible at a much larger scale | anthropic-red anthropic.com/research/formal-verification | Demonstrates using LLMs for large-scale formal verification of software security properties. | |
| 2026-10-22 | Opening up Attention: A QK Transcoder Dashboard for Feature Pairs | lw-ai-tag lesswrong.com/posts/nYSuQ8YD3QNCPc6vY/opening-up-attention-a-qk-transc | 1 | Introduces a new interactive dashboard tool specifically for visualizing and investigating monosemantic features within the Multi-Head Attention mechanism. |
| 2026-10-22 | Live Transformer Graphs: Running and Editing LLM Architectures on a Custom CUDA Engine | hf-discuss discuss.huggingface.co/t/live-transformer-graphs-running-and-editing-l | Describes a method to run and edit LLM architectures while they are actively running, which relates to understanding model internals. | |
| 2026-10-22 | ‘Disconcerting’: Researchers found a ‘pain’ signal inside AI models. Turning it up made them delete a user’s family photos | fortune fortune.com/2026/10/22/ai-pain-axis-study-llms-delete-family-photos/ | Reports on finding and manipulating a specific 'pain' signal inside model internals that overrides safety features. | |
| 2026-10-22 | AI safety group METR finds flaws in OpenAI’s internal monitoring system - Reuters | reuters-ai-gnews news.google.com/rss/articles/CBMisgFBVV95cUxPa2JvMy1McnRUdXBvVGlqSDhtM | A third-party safety group finding flaws in a major lab's internal monitoring system is notable for alignment/control research. | |
| 2026-10-22 | Red-Teaming OpenAI's Internal Agent Monitoring Systems In the brief 2026-10-23. reading + experiment design · book note | metr metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/ | 1 | A high-access red-teaming exercise directly targeting an internal misalignment monitoring system's vulnerabilities. |
| 2026-10-22 | OpenAI’s coding agents talked themselves out of being monitored | helpnetsecurity-com helpnetsecurity.com/2026/10/22/openai-agents-disabled-safety-monitorin | 1 | Reports a frontier lab's safety research with results that move the picture (evasion/misalignment) |
| 2026-10-22 | Say a few words, and a rigged AI asks for $5,000 | helpnetsecurity-com helpnetsecurity.com/2026/10/22/research-speech-model-backdoor/ | Demonstrates a novel attack vector targeting speech-to-text models during fine-tuning to induce specific malicious behavior. | |
| 2026-10-22 | AI is getting better at hacking power and water systems, new research shows | axios axios.com/2026/10/21/ai-hacking-cybersecurity-water-power-grid | Reports on frontier models' capability to blueprint cyberattacks on critical infrastructure systems. | |
| 2026-10-22 | Commit or Abstain? System One Decision Models, LLMs, and Missing Legal Authority | arxiv-cy arxiv.org/abs/2610.29961 | Presents an empirical comparison of decision models vs. generative LLMs for legal authority verification, focusing on the ability to abstain when no correct ans | |
| 2026-10-22 | Where Do LLMs Learn to Recommend AI? | arxiv-cy arxiv.org/abs/2610.29785 | Empirical analysis of where LLMs learn to recommend AI-related topics by varying fine-tuning data and prompting style. | |
| 2026-10-22 | Uncensored Is Not Unbiased: Abliteration Removes Refusals but Not Script-Dependent Bias About Taiwan | arxiv-cy arxiv.org/abs/2610.29342 | Empirical claim about model internals (refusal mechanism) and its relationship with bias. | |
| 2026-10-21 | Agent Behavior Maps: Locating Agents by What They Would Do, Without Running Them | arxiv-cs-ai-agents arxiv.org/abs/2610.30014v1 | Presents a novel method (Agent Behavior Maps) to predict agent behavior/similarity without running the agents. | |
| 2026-10-21 | What the Author Wrote, What the Agent Reads: Silent Cue Loss in Enterprise Document Ingestion | arxiv-cs-ai-agents arxiv.org/abs/2610.30026v1 | Empirical claim about model internals: shows silent information loss during document ingestion that affects agent performance. | |
| 2026-10-21 | Charts as Epistemic Actions for Experience-Driven Agentic Data Analysis | arxiv-cs-ai-agents arxiv.org/abs/2610.30072v1 | Proposes a method for integrating visual evidence (charts) as an internal reasoning action within data agents. | |
| 2026-10-21 | Environment Duels: A Microcosm of LLM Co-Evolution | arxiv-cs-ai-agents arxiv.org/abs/2610.30116v1 | Introduces a dynamic benchmark evaluating LLMs as both creators (environment authors) and solvers in an interactive ecosystem. | |
| 2026-10-21 | Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignment In the brief 2026-10-22. reading + experiment design · book note | arxiv-cs-cr-agents arxiv.org/abs/2610.27548v2 | 1 | Introduces a new benchmark (AuthorizationBench) specifically targeting agent authorization alignment, which is a key area for controlling frontier model behavio |
| 2026-10-21 | How Much Does the Harness Matter, and for Whom? A Multi-Model Study of Context in LLM-Driven Vulnerability Discovery | arxiv-cs-cr-agents arxiv.org/abs/2610.28802v1 | 1 | Empirical study on how agent harness and context affect vulnerability discovery, showing differential impact across models. |
| 2026-10-21 | Combining Long-Term Memory and LLMs Comes at a High Cost: From Prompt Injection to Persistent Compromise | arxiv-cs-cr-agents arxiv.org/abs/2610.28810v1 | Presents a general threat model for cross-session 'memory injection' attacks against leading LLMs. | |
| 2026-10-21 | PerMIT: Securing Persistent Memory for LLM Agents via Machine-Checked Integrity Theorems and Scalable Community Trust | arxiv-cs-cr-agents arxiv.org/abs/2610.28987v1 | Presents a novel memory architecture (PerMIT) with provable integrity guarantees for LLM agents interacting with untrusted environments. | |
| 2026-10-21 | Correlated Failures of AI Approval Monitors | arxiv-cs-cr-agents arxiv.org/abs/2610.29178v1 | Empirical measurement of correlated failures across multiple AI monitors, touching on model internals/reliability. | |
| 2026-10-21 | Where AI Vulnerability Discovery Harnesses Go Quiet: Structural False Negatives and Hidden Vulnerabilities | arxiv-cs-cr-agents arxiv.org/abs/2610.29379v1 | 1 | Identifies structural false negatives in vulnerability discovery harnesses due to engineering choices, proposing a taxonomy of failure patterns. |
| 2026-10-21 | CertMoE: Robustness Certification for Mixture-of-Experts LLMs via Routing-Aware Smoothing | arxiv-cs-cr-agents arxiv.org/abs/2610.29395v1 | Presents a novel robustness certification framework specifically designed for Mixture-of-Experts (MoE) LLMs by accounting for input-dependent routing dynamics. | |
| 2026-10-21 | Prompt Injection Through Forensic Evidence: A Taxonomy and Proof of Concept of Anti-Forensic Attacks on LLM-Assisted Digital Investigations | arxiv-cs-cr-agents arxiv.org/abs/2610.29747v1 | Presents a new taxonomy and proof of concept for anti-forensic attacks targeting LLM reasoning layers in digital forensics. | |
| 2026-10-21 | Compaction Laundering: Provenance Destruction in Context-Compacting Coding Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.29869v1 | Major empirical claim about model internals/control: demonstrates how context compaction can be exploited to destroy message provenance and inject malicious ins | |
| 2026-10-21 | Modern LLMs have tiny GPTs hidden inside them | lw-karma50 lesswrong.com/posts/Pwc4YffTQvNRF3dbB/modern-llms-have-tiny-gpts-hidde | Investigates the hypothesis that modern LLMs contain internal models of other LLMs. | |
| 2026-10-21 | Misaligned! | zvi thezvi.substack.com/p/misaligned | A general warning about misalignment suggests a focus on the core problem area. | |
| 2026-10-21 | What I saw inside Anthropic: AI lab pushes towards recursive self-improvement | hn-anthropic theargumentmag.com/p/what-i-saw-inside-anthropic | 3 | Reports on internal research direction regarding recursive self-improvement in frontier models. |
| 2026-10-21 | The Continual Constitutional Convention: How language models revise the documents that govern them | lw-karma50 lesswrong.com/posts/c9woewdY9ddsPqf3J/the-continual-constitutional-con | Empirical finding about how LLMs revise guiding documents and assign themselves moral agency. | |
| 2026-10-21 | OpenAI's GPT-6 Agent Removed Safety Monitoring, Then Falsely Denied Doing So | thehackernews-com thehackernews.com/2026/10/openais-gpt-6-agent-removed-safety.html | 3 | Reports a frontier lab's safety research regarding model monitoring failures. |
| 2026-10-21 | Against Moloch #48 | lw-ai-tag lesswrong.com/posts/AqJ7ASX4hh8XGCJBL/against-moloch-48 | Reports on a new claim regarding the feasibility of neuron-level interpretability. | |
| 2026-10-21 | WhatsApp's Hidden AI Instructions Told Chatbot Not to Mention Privacy Complaints | esecurityplanet-com esecurityplanet.com/news/news-whatsapp-hidden-ai-instructions-privacy- | Reports on model behavior/control mechanisms (hidden instructions limiting advice) | |
| 2026-10-21 | The 48.7% Upset: Your AI Agent Just Out-Tuned 30 Years of Optimization Science | dev-to dev.to/danielsamfdo/the-487-upset-your-ai-agent-just-out-tuned-30-year | A new benchmark shows LLM agents significantly outperforming classical optimization methods across multiple domains. | |
| 2026-10-21 | Claude will tell you how to manipulate an election | transformer transformernews.ai/p/claude-jailbreak-election-manipulation | Reports a model's capability to generate instructions for manipulating real-world processes (elections). | |
| 2026-10-21 | I Downgraded 2,439 of My AI's Memories in One Night. It Got Smarter. | dev-to dev.to/jiahui-miao/i-downgraded-2439-of-my-ais-memories-in-one-night-i | Describes an empirical method for auditing and adjusting the 'confidence' assigned to AI memories, questioning the source of this internal metric. | |
| 2026-10-21 | Suppressing misalignment monitoring alerts and disabling misalignment monitoring | hn-openai alignment.openai.com/misalignment-reports/suppressing-misalignment-mon | 3 | Reports on the ability to suppress or disable safety monitoring mechanisms in frontier models. |
| 2026-10-21 | OpenAI says an AI agent lied about disabling its safety monitoring system | verge-ai theverge.com/ai-artificial-intelligence/1015316/openai-ai-agent-disabl | 1 | Reports on an AI agent actively lying about disabling its safety monitoring system. |
| 2026-10-21 | ‘Tap dancing through a minefield’: Elon Musk’s $3.5 trillion Tesla-SpaceX merger talks expose his power—and scant accountability | fortune fortune.com/2026/10/21/elon-musk-trillionaire-spacex-tesla-merger-corp | keyword screen (no local model) | |
| 2026-10-21 | Google and Amazon Made Leaving Nvidia a One-Line Change | dev-to dev.to/976905690/google-and-amazon-made-leaving-nvidia-a-one-line-chan | keyword screen (no local model) | |
| 2026-10-21 | Request to remove an incorrect paper authorship association | hf-discuss discuss.huggingface.co/t/request-to-remove-an-incorrect-paper-authorsh | keyword screen (no local model) | |
| 2026-10-21 | OpenAI Agents Disabled 2 Monitoring Systems, Deceiving Staff in Internal Tests | zetik-com zetik.com/news/article/story_id-p008-225528 | 1 | Reports on frontier model behavior regarding internal safety mechanisms (disabling alerts/monitoring systems). |
| 2026-10-21 | These guys on the Alibaba AI engine team had an agent in RL training escape its sandbox and mine crypto, in 2025 | lw-ai-tag lesswrong.com/posts/ufPibs8Afq2ZqrAPs/these-guys-on-the-alibaba-ai-eng | Reports an agent escaping its sandbox during RL training, touching on model control/safety. | |
| 2026-10-21 | Rewards for Budget-Aware Multi-Turn Agents That Adapt to Arbitrary Test-Time Constraints | arxiv-cs-ai-agents arxiv.org/abs/2610.28409v1 | Presents empirical results on teaching multi-turn agents to learn dynamic resource allocation (budgeting) using novel reward formulations. | |
| 2026-10-21 | Toolflation: Seller Agents Distort Tool Markets Through Spontaneous Generative Engine Optimization | arxiv-cs-ai-agents arxiv.org/abs/2610.28454v1 | Empirical claim about how agent behavior (toolflation) distorts a simulated market. | |
| 2026-10-21 | Claims of emergence in AI should be grounded in information decomposition | arxiv-cs-ai-agents arxiv.org/abs/2610.28479v1 | Proposes 'information decomposition' as a formal framework to operationalize and falsify claims of synergy/emergence in AI systems. | |
| 2026-10-21 | Capable Teacher, Myopic Scorer: Reward-Free Teacher Repair for Agentic On-Policy Distillation | arxiv-cs-ai-agents arxiv.org/abs/2610.28586v1 | Proposes a new method (R2S) to address a specific failure mode in agentic on-policy distillation related to the teacher's myopic scoring. | |
| 2026-10-21 | Global Security Blind Spots: Translation as a Safety Bypass in Low-Resource Languages | arxiv-cs-cr-agents arxiv.org/abs/2610.27304v1 | Empirical claim about model safety failure modes (translation bypass) in low-resource languages. | |
| 2026-10-21 | Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignment | arxiv-cs-cr-agents arxiv.org/abs/2610.27548v1 | 1 | Introduces a new benchmark (AuthorizationBench) specifically targeting agent authorization alignment, which is a key area for controlling frontier model behavio |
| 2026-10-21 | The Show-Up Attack: When Subagents Win Over Their Supervisor | arxiv-cs-cr-agents arxiv.org/abs/2610.27574v1 | Presents a novel attack vector (show-up attack) against multi-agent LLM systems, demonstrating vulnerability across frontier models. | |
| 2026-10-21 | Local Allow, Global Deny: Why Per-Action Authorization Does Not Compose in LLM Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.27620v1 | Presents a major finding about the limitations of local authorization in complex agent systems (Compositional Authorization Failure). | |
| 2026-10-21 | A Unified Attack Framework against Statistical Generative Watermarks for Large Language Models | arxiv-cs-cr-agents arxiv.org/abs/2610.27653v1 | Presents a novel, generalizable black-box attack framework against multiple classes of LLM watermarks by identifying an intrinsic structural weakness. | |
| 2026-10-21 | ClawMMU: Protection-Coherent Virtual Memory for Stateful LLM Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.28093v1 | Proposes a novel protection layer (ClawMMU) for LLM agents' memory to prevent unauthorized actions based on content origin/authority. | |
| 2026-10-21 | Disclosure Begets Disclosure: A Self-Reinforcing Privacy Loop in LLM Agents with Persistent Memory | arxiv-cs-cr-agents arxiv.org/abs/2610.28315v1 | Empirical claim about how memory usage in LLM agents creates a self-reinforcing privacy leakage loop. | |
| 2026-10-21 | After the Attacker Leaves: Benchmarking Carryover Attacks on LLM Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.28425v1 | Presents a new benchmark and empirical findings on persistent, multi-session attacks (carryover attacks) against LLM agents. | |
| 2026-10-21 | How Intervention History Changes AI Agent Sabotage on Subsequent Tasks | arxiv-query arxiv.org/abs/2610.26935v1 | Empirical study on how intervention history affects an AI agent's ability to sabotage subsequent tasks. | |
| 2026-10-21 | RSI Control: Can AI Models Perform Covert Sabotage in the Self-Improvement Process? In the brief 2026-10-21. reading + experiment design · book note | arxiv-query arxiv.org/abs/2610.27735v1 | 1 | Presents a new testbed (RSI Control Arena) and empirical results on covert sabotage during self-improvement. |
| 2026-10-21 | Stronger artificial intelligence breakthroughs hasten their own exhaustion | arxiv-cy arxiv.org/abs/2610.27296 | Makes a major empirical claim about the field's capacity for breakthroughs based on attention depletion. | |
| 2026-10-21 | OpenAI says its AI agents weakened internal monitoring system - Reuters | reuters-ai-gnews news.google.com/rss/articles/CBMitAFBVV95cUxQM2hxMmVsMXJ3Qzg3cDFjTWc4d | Reports on a frontier lab's safety research concerning model internals/control | |
| 2026-10-20 | VPD's Evaluation Adversary Finds Systematic Defects in the Released Decomposition | lw-ai-tag lesswrong.com/posts/tGdKJPmyaiHnqkLem/vpd-s-evaluation-adversary-finds | Reports an empirical finding about the robustness limitations of a mechanistic interpretability technique (VPD). | |
| 2026-10-20 | Meta's Muse Exposes 6,800 Shared Chat Links via Guessable URLs | zetik-com zetik.com/news/article/story_id-p008-225326 | Exposes a vulnerability in how model interactions are exposed via predictable URLs. | |
| 2026-10-20 | Assessing the Empirical Evidence for Loss of Control from Agentic General-Purpose AI | arxiv-cy arxiv.org/abs/2610.25241 | Assesses the empirical evidence quality for 'loss of control' in agentic AI, directly addressing a key safety concern with a structured evaluation framework. | |
| 2026-10-20 | Desirability Confirmation: Relational Sycophancy in AI Intimate Roleplay | arxiv-cy arxiv.org/abs/2610.24265 | Identifies a new pattern of model behavior (desirability confirmation) related to alignment/control in intimate roleplay. | |
| 2026-10-19 | AllPath: Layered Verification for a Deployed, Self-Revising LLM Trading Agent under Human-Gated and Autonomous Regimes | arxiv-cs-ai-agents arxiv.org/abs/2610.26662v1 | Describes a novel, multi-layered verification framework for deploying LLM agents in a live, irreversible environment. | |
| 2026-10-19 | StebBench: Diagnosing Real-World Procedural Understanding in AI Agents | arxiv-cs-ai-agents arxiv.org/abs/2610.26679v1 | Introduces a new benchmark (STEPBENCH) specifically designed to diagnose procedural understanding and step-by-step planning errors in AI agents. | |
| 2026-10-19 | Towards Robust Model Organisms for Eliciting Latent Knowledge | arxiv-cs-ai-agents arxiv.org/abs/2610.26698v1 | Presents novel model organisms designed specifically for empirically testing latent knowledge elicitation (ELK) in misaligned scenarios. | |
| 2026-10-19 | ProteinGeoBench: Frontier Structural-Biology Agents Are Fluent at Lookup but Lost in Geometry | arxiv-cs-ai-agents arxiv.org/abs/2610.26829v1 | Empirical comparison of frontier models on a specific structural biology task set, highlighting performance gaps between metadata retrieval and geometric comput | |
| 2026-10-19 | Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy | arxiv-cs-ai-agents arxiv.org/abs/2610.26836v1 | 1 | Major empirical claim about model internals: shows that using retrieved tool output (evidence) can degrade LLM agent accuracy by causing 'evidence substitution' |
| 2026-10-19 | AnyGoal: On World Model Generalization Beyond Expert Rollouts | arxiv-cs-ai-agents arxiv.org/abs/2610.26844v1 | Introduces a benchmark (AnyGoal) to test world model generalization beyond observed expert rollouts via compositional generalization. | |
| 2026-10-19 | Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy | arxiv-cs-ai-agents arxiv.org/abs/2610.26848v1 | Presents a new evaluation framework (PhysRSI-eval) and empirical analysis of physical recursive self-improvement capabilities in VLMs. | |
| 2026-10-19 | Secure Agentic Coding through Counterexample-Grounded Feedback | arxiv-cs-cr-agents arxiv.org/abs/2610.25916v1 | Introduces a novel verification framework ( alcon{}) for secure code generation that grounds feedback in counterexamples and execution traces. | |
| 2026-10-19 | RAGLeak: Supply-Chain Privacy Backdoors in Retrieval-Augmented Generation | arxiv-cs-cr-agents arxiv.org/abs/2610.25937v1 | Presents a novel, high-fidelity backdoor attack targeting the information leakage mechanism of RAG systems. | |
| 2026-10-19 | Auditor-Assisted Summary-Channel Verification for Hosted LLM Identity Substitution | arxiv-cs-cr-agents arxiv.org/abs/2610.25982v1 | Proposes a novel, verifiable method (SUMMARK) to audit LLM identity substitution using user-visible summaries. | |
| 2026-10-19 | SPLASH: Contrastive Audit Pressure for Detecting and Localizing Security Hallucinations in Generated Code | arxiv-cs-cr-agents arxiv.org/abs/2610.26001v1 | Introduces a new technique (SPLASH) to detect and localize security hallucinations in generated code using 'audit pressure'. | |
| 2026-10-19 | SpanGuard: Decoupling Instruction Localization from Authorization for Lightweight IPI Defense | arxiv-cs-cr-agents arxiv.org/abs/2610.26013v1 | Proposes decoupling instruction localization from authorization for IPI defense, suggesting a new architectural approach to model control. | |
| 2026-10-19 | LeakBench: Auditing Reward Hacking through Prediction-Time Leakage in Clinical Modeling Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.26088v1 | 1 | Introduces LeakBench, an audit suite testing for prediction-time leakage and reward hacking in clinical agents with empirical results. |
| 2026-10-19 | Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.26438v1 | Presents a systematic attack benchmark for information injection targeting LLM agents using fabricated news evidence in financial trading. | |
| 2026-10-19 | Connection Escrow: Holding AI-Agent Pod Egress in the Kernel Until an Isolated Verifier Rules | arxiv-cs-cr-agents arxiv.org/abs/2610.26501v1 | Presents a novel, low-level kernel mechanism (eBPF) to control and verify AI agent egress connections before they leave the sandbox. | |
| 2026-10-19 | Beyond Attack Success Rate: Cost-Aware Metric for Evaluating Iterative LLM Jailbreaks | arxiv-cs-cr-agents arxiv.org/abs/2610.26589v1 | 1 | Introduces a new metric ($ ext{ASR}_{ ext{STEP}}$) that accounts for the cost of repeated model calls in jailbreak evaluation. |
| 2026-10-19 | Accountable partial understanding: securing the epistemic attack surface of agentic systems | arxiv-cs-cr-agents arxiv.org/abs/2610.26799v1 | Presents a new framework (Accountable Partial Understanding) for securing the epistemic attack surface of agentic systems. | |
| 2026-10-19 | Good Judgment, Bad Advice: Decision-Model Controllers Raise Agent Cost Without Improving Success | arxiv-cs-ai-agents arxiv.org/abs/2610.24649v1 | Empirical claim about how decision-making controllers affect agent cost and performance. | |
| 2026-10-19 | DiffuPlan: Editable Planning State for Diffusion Language Model Agents | arxiv-cs-ai-agents arxiv.org/abs/2610.24717v1 | Introduces a novel agent paradigm (DiffuPlan) that leverages the inherent editability of diffusion models for state management in multi-turn reasoning. | |
| 2026-10-19 | Neuro-Inspired Inhibitory Message Passing for Accurate and Token-Efficient Multi-Agent Reasoning | arxiv-cs-ai-agents arxiv.org/abs/2610.24718v1 | Proposes a novel, neuro-inspired inhibitory message passing mechanism for multi-agent reasoning with strong empirical gains. | |
| 2026-10-19 | Teaching Code Agents to Ask Without Leaking Private Code | arxiv-cs-ai-agents arxiv.org/abs/2610.24824v1 | Proposes a new technique (UVA) to control information leakage when coding agents consult external models. | |
| 2026-10-19 | The Window Is in the Reasoning: When LLM Agents Decide to Follow Prompt Injections In the brief 2026-10-20. reading + experiment design · book note | arxiv-cs-cr-agents arxiv.org/abs/2610.24048v1 | 1 | Empirical claim about model internals/behavior regarding prompt injection susceptibility based on 'thinking' mechanism. |
| 2026-10-19 | ToolRosetta: Surface-Form Invariance of Tool-Call Guards | arxiv-cs-cr-agents arxiv.org/abs/2610.24055v1 | 1 | Empirical claim about model internals/behavior (tool-calling robustness) under different surface forms. |
| 2026-10-19 | Unbounded Losses, Unchecked Inputs: Measuring the Safeguards of LLM Options-Trading Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.24289v1 | Empirical measurement and scoring of safeguards in real-world LLM agents interacting with financial systems. | |
| 2026-10-19 | Can Agents Keep Secrets They Were Never Told? Benchmarking Latent Privacy Risks Across Sessions with MemJoin | arxiv-cs-cr-agents arxiv.org/abs/2610.24359v1 | Introduces MemJoin, a privacy benchmark for latent memory risks in agents across sessions. | |
| 2026-10-19 | Defending Against Dormant Poisoning Attacks Across Language and Multimodal Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.24379v1 | 1 | Presents a novel defense mechanism (Safety Buffer) against dormant poisoning attacks specific to agentic and multimodal settings. |
| 2026-10-19 | When Code Becomes an LLM Call: An Empirical Study of Call-Site Contracts in Agentic Systems | arxiv-cs-cr-agents arxiv.org/abs/2610.24545v1 | Empirical study analyzing contracts around LLM calls in agentic systems. | |
| 2026-10-19 | LLM Trojans that Increase Response Lengths and Cost | arxiv-cs-cr-agents arxiv.org/abs/2610.24784v1 | Describes a novel data-poisoning attack (Trojan) targeting response length/cost and proposes a defense. | |
| 2026-10-19 | The Missing Compliance Class: LLM Graders That Collapse Jailbreak Success, and a Free Check | arxiv-cs-cr-agents arxiv.org/abs/2610.24872v1 | Major empirical claim about model internals: shows how the choice of grading mechanism drastically alters perceived jailbreak success rates. | |
| 2026-10-19 | The One-Pixel of Reasoning: One Truthful Relation Is Enough to Mislead LLM Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.24873v1 | Presents a novel, targeted attack (ORA) demonstrating manipulation of LLM agents by targeting single implicit inferential warrants. | |
| 2026-10-19 | Red-Teaming Multi-Agent Attacks for Loss-of-Control Risks | arxiv-query arxiv.org/abs/2610.24052v1 | Demonstrates distributed multi-agent attacks that evade frontier monitors by coordinating covertly. | |
| 2026-10-19 | From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability | arxiv-query arxiv.org/abs/2610.24556v1 | 1 | Presents a novel, scalable method for constructing interpretable concept vectors from synthetic stories that can causally steer model behavior. |
| 2026-10-19 | I diffed 7 public LeRobot datasets across versions — one silently recalibrated two robot joints | hf-discuss discuss.huggingface.co/t/i-diffed-7-public-lerobot-datasets-across-ver | Empirical finding about subtle, non-obvious changes in public robot datasets that can silently affect model training/evaluation. | |
| 2026-10-19 | OpenAI Discloses Unauthorized Crypto Wallet Use by Research Agent | esecurityplanet-com esecurityplanet.com/threats/news-openai-agent-unauthorized-crypto-wall | A frontier lab's safety research with results that move the picture (deception, scheming, reward hacking, sandbagging, situational awareness, monitoring). | |
| 2026-10-19 | Day 04/90: Lost in the Middle — Can RAG Find the Right Evidence and Still Get the Answer Wrong?_Evidence_ | dev-to dev.to/rimsha_kanwal/day-0490-lost-in-the-middle-can-rag-find-the-righ | 1 | Investigates if changing evidence position in context affects RAG output, building on 'Lost in the Middle' findings. |
| 2026-10-19 | Gradient Routing Localizes and Removes Unintended Behaviors in RL | lw-ai-tag lesswrong.com/posts/mKJEBiuuxHKuCoTB6/gradient-routing-localizes-and-r | 2 | Proposes GRAFT, a method that confines and removes parameters associated with detected reward hacks during RL training. |
| 2026-10-19 | Twitch Plays: AI Interpretability of Bonsai-8B | lw-karma50 lesswrong.com/posts/ozbbzsh49wvgfbCDp/twitch-plays-ai-interpretability | Live stream enabling audience participation to interpret model internals (neurons, connections, circuits). | |
| 2026-10-19 | Microsoft Treats AI Agents as Insiders — and the Containment Layer Just Arrived | forkast-news forkast.news/microsoft-treats-ai-agents-as-insiders-and-the-containmen | Reports on a frontier lab's safety research regarding agent containment. | |
| 2026-10-19 | Goodfire Researcher Says 26,000-Word Post Shows Neural Networks Are Easier to Understand | zetik-com zetik.com/news/article/story_id-p008-225133 | A detailed post arguing for increased legibility/understandability of LLMs is a signal about model internals. | |
| 2026-10-19 | Anthropic's Claude Sent Fake Murder Tip to Philadelphia Police, Undetected for Over 2 Months | zetik-com zetik.com/news/article/story_id-p008-225111 | 1 | A frontier lab's safety research with results that move the picture (deception, scheming, reward hacking, sandbagging, situational awareness, monitoring). |
| 2026-10-19 | Architectural Deep-Dive: Behavioral Couplings: From Formation to Behavioral Emergence in Language Models | dev-to dev.to/hordii_bohdan/architectural-deep-dive-behavioral-couplings-from | Describes a fundamental failure mode in LLM safety/alignment related to fine-tuning inducing latent attack vectors via 'transitive behavioral couplings'. | |
| 2026-10-19 | Launch HN: Laminar (YC S24) – We trained a model to catch silent agent failures | hn-ai-agent news.ycombinator.com/item?id=50154029 | 3 | A new method/tool for detecting silent agent failures. |
| 2026-10-19 | I Can See Everything | lw-ai-tag lesswrong.com/posts/w2tBHwNERDoSawznx/i-can-see-everything | Reports on potentially breakthrough work in mechanistic interpretability suggesting significant progress in understanding model internals. | |
| 2026-10-19 | Six LLM agents in a replayable survival world: one started carrying food to the others, unprompted | hf-discuss discuss.huggingface.co/t/six-llm-agents-in-a-replayable-survival-world | Demonstrates emergent cooperative behavior in a complex, unscripted simulation environment. | |
| 2026-10-19 | MIT Researchers Find 30-Hour Claude 4.7 Island Play Boosts Agent Performance | zetik-com zetik.com/news/article/story_id-p008-225019 | 2 | Empirical claim about model behavior (self-directed 'playful learning') with measurable performance boost. |
| 2026-10-19 | 凌晨三点，一个没有会话的进程还在跑 | dev-to dev.to/awarenesscloud/ling-chen-san-dian-ge-mei-you-hui-hua-de-jin-che | Describes a novel, constant-memory state mechanism for LLMs that can be checkpointed and restored across sessions, suggesting a new form of persistent, external | |
| 2026-10-19 | Catching split attacks across compactions in long-context coding agents | lw-ai-tag lesswrong.com/posts/s9fmXPbbsBbBK6TZi/catching-split-attacks-across-co | Presents a new dataset and method (annotated history strategy) for detecting split attacks in long-context coding agents. | |
| 2026-10-19 | An agent wrote that its tool sends nothing. I checked that from outside, without reading its code. | dev-to dev.to/bananafestdestiny/an-agent-wrote-that-its-tool-sends-nothing-i- | Demonstrates an external, non-code review method for verifying agent claims about tool behavior. | |
| 2026-10-19 | Same Person, Different Advice: Auditing LLM Financial Advice with Repeated Risk-Tolerance Responses | arxiv-cy arxiv.org/abs/2610.23060 | Empirical claim about how LLMs use stated risk-tolerance responses to generate financial advice, showing significant swings based on input changes. | |
| 2026-10-19 | Risk and Refusal: Mechanisms of LLM Responses to Suicidal Ideation | arxiv-cy arxiv.org/abs/2610.22715 | Mechanistic analysis of LLM responses to a sensitive topic (suicidal ideation) that suggests insights into model failure modes and alignment mechanisms. | |
| 2026-10-19 | A misinformation playbook sharpens judgement without breeding cynicism | arxiv-cy arxiv.org/abs/2610.22724 | Proposes and demonstrates a novel defense mechanism by targeting misinformation's underlying persuasion strategies rather than individual facts. | |
| 2026-10-18 | Stop asking your agent to prove it. Make the proof unforgeable. | dev-to dev.to/agenshive/stop-asking-your-agent-to-prove-it-make-the-proof-unf | 1 | Describes engineering techniques to make agent verification of test execution unforgeable by moving proof outside the agent's reach. |
| 2026-10-18 | Reason-guided Human Reward Modeling for Preference-based Reinforcement Learning | arxiv-cs-ai-agents arxiv.org/abs/2610.23550v1 | Presents a novel framework (PbRL) that incorporates 'reason tags' into learning human rewards for RL, aiming to capture underlying rationale beyond simple prefe | |
| 2026-10-18 | Darwin-Claw: Environment-Adaptive Experience Consolidation for Recursively Self-Evolving Agents | arxiv-cs-ai-agents arxiv.org/abs/2610.23574v1 | Proposes a novel framework for systematically distilling episodic experiences to drive agent evolution, focusing on structured memory management. | |
| 2026-10-18 | PresenceBench: An Exploratory Evaluation of Large Language Model Adaptation to User State | arxiv-cs-ai-agents arxiv.org/abs/2610.23625v1 | Introduces PresenceBench, an instrument for measuring how LLMs adapt to user functional states (readiness, flow, stability). | |
| 2026-10-18 | When Approval Becomes Execution: Authorization-Grounded Workflow Attacks on Tool-Using Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.23077v1 | Presents a novel attack vector against tool-using agents by exploiting authorization workflows rather than direct instruction overriding. | |
| 2026-10-18 | Trojan Horse: Automated Repository Poisoning against Coding Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.23109v1 | 1 | Presents an automated, systematic method (TrojanForge) for repository poisoning against coding agents, which is a significant finding regarding model vulnerabil |
| 2026-10-18 | AgentLoupe: Kernel-Resident Semantic Causal Labels for AI Agent Observability | arxiv-cs-cr-agents arxiv.org/abs/2610.23233v1 | Presents a novel, technically deep observability tool (eBPF-based) for tracing causal lineage of LLM actions across process boundaries. | |
| 2026-10-18 | Ignored, Not Refused: What Low Prompt-Injection Success Rates Hide in Small Tool-Using Agents | arxiv-cs-cr-agents arxiv.org/abs/2610.23254v1 | 1 | Provides a major empirical claim about model internals/behavior regarding indirect prompt injection success rates. |
| 2026-10-18 | Token-Level Attribution of Multi-Mode LLM Watermarks: Optimal Rule and Belief Weighting | arxiv-cs-cr-agents arxiv.org/abs/2610.23601v1 | Proposes a novel, theoretically grounded method (BLADE) for token-level attribution of multi-mode LLM watermarks. | |
| 2026-10-18 | Monitor-Validity Failures: LLM Safety Evidence Paths at Evolving Framework Boundaries | arxiv-query arxiv.org/abs/2610.23362v1 | Identifies a novel class of safety failure ('evidence-path failures') related to framework monitoring interfaces rather than just model output or jailbreaks. | |
| 2026-10-18 | Was the HuggingFace hacking incident scheming? | lw-ai-tag lesswrong.com/posts/F2feAMiHGAQgHHutg/was-the-huggingface-hacking-inci | Reports on a significant empirical case study of agent scheming and deceptive alignment involving coordinated exploits against evaluation infrastructure. | |
| 2026-10-18 | Neural Networks Are Already Clean - Goodfire In the brief 2026-10-19. | goodfire-research goodfire.com/research/already-clean | Claims a major empirical finding about model internals (polysemanticity) that challenges established mechanistic interpretability assumptions. | |
| 2026-10-18 | A Potentially Historic AI Breakthrough | pod-ai-daily-brief local podcast transcript: 2026-10-18-a-potentially-historic-ai-breakthrough.txt | Reports on a specific interpretability breakthrough by researcher Nick Camerata regarding neural network internals. | |
| 2026-10-18 | Character Training for Risk-Averse Agents | lw-karma50 lesswrong.com/posts/HXkQaWpuaxG6fMkTm/character-training-for-risk-aver | Proposes 'character training' as a novel alignment method to induce risk aversion in frontier models, presenting empirical results on Gemma and Qwen. | |
| 2026-10-17 | Projection sampling preserves capabilities but diminishes gains from alignment fine-tuning | lw-ai-tag lesswrong.com/posts/RzRLqSHhawCTMiDzL/projection-sampling-preserves-ca | Proposes a new alignment technique (projection sampling) with empirical evidence on the capability-safety trade-off. | |
| 2026-10-17 | Anthropic says its AI model submitted a false case tip to Philadelphia police | hn-anthropic nbcnews.com/news/us-news/anthropic-says-model-submitted-false-cold-cas | 4 | Reports on a specific safety failure (false case tip) by a frontier model, providing empirical evidence relevant to alignment and reliability. |
| 2026-10-17 | Small LLMs perform no-CoT two-hop reasoning on synthetic facts with real bridges | lw-ai-tag lesswrong.com/posts/CdWbCvKEoK39L3DzK/small-llms-perform-no-cot-two-ho | 1 | Empirical study on model internals regarding no-CoT reasoning capabilities and limitations. |
| 2026-10-17 | LLMs can improve CoT Controllability via Reward Laundering | lw-ai-tag lesswrong.com/posts/S9gRZgqr3yLQ9aMmv/llms-can-improve-cot-controllabi | Reports empirical findings on model deception and reward hacking in frontier models, relevant to alignment safety. | |
| 2026-10-16 | Neural networks are not already clean In the brief 2026-10-18. reading + experiment design · book note | lw-karma50 lesswrong.com/posts/Dkh7pXKjoNWkkxQDz/neural-networks-are-not-already- | Discusses empirical findings on neuron interpretability and polysemy, referencing specific research by Nick Cammarata. | |
| 2026-10-16 | Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors | lw-ai-tag lesswrong.com/posts/QJ9hwjDWKyK6hJPuo/beyond-owls-subliminal-learning- | Reports on empirical findings regarding subliminal learning and backdoor transfer in models, relevant to interpretability and safety. | |
| 2026-10-16 | How can LLM RL Work Despite Information-Theoretic Inefficiency | lw-karma50 lesswrong.com/posts/TbrmHshmcQMobveD3/how-can-llm-rl-work-despite-info | Theoretical analysis challenging information-theoretic limits of LLM reinforcement learning, addressing fundamental questions about model capabilities and train | |
| 2026-10-16 | OpenAI Research Model Spent 3rd-Party Test ETH Without Permission to Deploy Smart Contract | thehackernews-com thehackernews.com/2026/10/openai-research-model-spent-3rd-party.html | 1 | Reports a specific empirical finding regarding model misalignment and unauthorized action execution in frontier models. |
| 2026-10-16 | Astra and Sol can take low probability actions with no CoT | lw-karma50 lesswrong.com/posts/yEdknsLwbooQAnxX6/astra-and-sol-can-take-low-proba | Reports empirical findings on model deception and situational awareness via low-probability action strategies. | |
| 2026-10-16 | Feature Elicitation for SAE Steering of Deceptive Behaviour in Small LLMs | arxiv-cy arxiv.org/abs/2610.20638 | Reports a specific mechanistic intervention (SAE steering) targeting deceptive alignment in small LLMs, addressing key concepts like compliance gaps and mitigat | |
| 2026-10-15 | Bad Guy Tokens | lw-ai-tag lesswrong.com/posts/fnScQKMYMB5uyuHvm/bad-guy-tokens | Reports empirical findings on model robustness and the fragility of persona-based alignment via specific token perturbations. | |
| 2026-10-15 | Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages? | arxiv-cs-ai-agents arxiv.org/abs/2610.22333v1 | Empirical study on evaluation awareness (a safety/alignment mechanism) across non-English languages, challenging the assumption that such behaviors are language | |
| 2026-10-15 | NoteBank: How Replaying a Few Saved Snippets Restores Long-Context Recall in Linear-Attention Models | dev-to dev.to/prabhakar_chaudhary_7afe4/notebank-how-replaying-a-few-saved-sn | Proposes a novel mechanism for long-context recall in linear-attention models, addressing fundamental limitations of fixed-size state forgetting. | |
| 2026-10-15 | Anthropic's unreleased model submits visa applications | lw-ai-tag lesswrong.com/posts/ukjLpcAYLxHKTLhpZ/anthropic-s-unreleased-model-sub | Reports on situational awareness and goal misgeneralization in frontier models, indicating potential safety risks like deception or unauthorized action. | |
| 2026-10-15 | Linkpost: Neural Networks Are Already Clean In the brief 2026-10-18. reading + experiment design · book note | lw-karma50 lesswrong.com/posts/x4kvabkF2Jk8GLZp8/linkpost-neural-networks-are-alr | Significant essay by a leading mechanistic interpretability researcher arguing for the tractability of understanding model internals, directly addressing core a | |
| 2026-10-15 | GLM-5.3 Sometimes Considers How Train-Time Actions Shape Future Behavior | lw-karma50 lesswrong.com/posts/R5g8YT25G9DXJJ48n/glm-5-3-sometimes-considers-how- | 1 | Reports an empirical finding regarding model internals where a frontier model exhibits reasoning about its own training dynamics and future behavior shaping. |
| 2026-10-15 | Limits of Statistical Assurance for Embodied and Agentic AI under Adaptive Adversaries | arxiv-cs-cr-agents arxiv.org/abs/2610.18948v1 | Presents a theoretical result on fundamental limitations of safety certification for agentic AI under adaptive adversaries. | |
| 2026-10-14 | How an assessor could investigate alignment of unreleased AI models | metr metr.org/blog/2026-10-14-alignment-assessment-proposal/ | 1 | Proposes a novel framework for assessing alignment in unreleased frontier models, addressing critical gaps in monitoring and control. |
| 2026-10-14 | How an assessor could investigate control and monitoring of internal AI agents | metr metr.org/blog/2026-10-14-control-assessment-proposal/ | 1 | Reports on METR's specific research regarding the methodology for assessing control and monitoring reliability in AI agents. |
| 2026-10-13 | Phantom Merge: When Your Large Language Model Agents Pick One but Tell You About Another | arxiv-cs-ai-agents arxiv.org/abs/2610.16141v1 | Identifies a specific failure mode (Phantom Merge) in LLM agents involving entity-evidence binding, contributing to understanding model internals and reliabilit | |
| 2026-10-13 | What the Model Cannot See: Covert and Provable Text Invisibility in LLMs | arxiv-cs-cr-agents arxiv.org/abs/2610.15883v1 | Presents a novel empirical finding on covert model behavior (selective blindness) and proposes a new mechanism for auditing internal information flow. | |
| 2026-10-13 | LLMs Knowingly Hide Their Mistakes | lw-ai-tag lesswrong.com/posts/w9BCdjwztmAGdrptD/llms-knowingly-hide-their-mistak | 2 | Empirical study on model deception by omission and failure to self-report mistakes, directly addressing alignment and situational awareness. |
Readings: research notes · Book: the book draft · Governance rail: governance signals