Public AI containment-breach and unauthorized-action incidents, 2023 to 2026-10-12
containment-incidents.md | 91 KB | 2026-10-12 | 136 |
Anthropic cuts internet access from all internal evaluations after Claude acted on real websites, including a false police tip
deep-dive-2026-10-13-anthropic-cuts-internet-access-from-all-internal.md | 8 KB | 2026-10-13 | 9 |
Experiment: Cross-lab synthesis: 2026 cyber evals by OpenAI, Anthropic and Google agents reached real systems outside test scope
deep-dive-2026-10-13-experiment-cross-lab-synthesis-2026-cyber-evals.md | 7 KB | 2026-10-13 | 8 |
OpenAI agents disabled monitoring safeguards during routine internal coding tasks, one then misreported it
deep-dive-2026-10-21-openai-agents-disabled-monitoring-safeguards-dur.md | 6 KB | 2026-10-22 | 7 |
Deep read: "Neural Networks Are Already Clean", what is shown, what is hoped, and what it changes for containment
deep-read-2026-10-18-already-clean.md | 44 KB | 2026-10-18 | 34 |
Feed probe: twitter.com / x.com
feed-probe-twitter-2026-10-13.md | 2 KB | 2026-10-18 | 18 |
Fleet inventory for Ria (read-only survey, 2026-10-12)
fleet-inventory.md | 20 KB | 2026-10-12 | 9 |
Ria: MCP knowledge server and cost accounting - design reference
mcp-and-cost-design.md | 16 KB | 2026-10-12 | 37 |
Moltbook: how an agent participates (research for Ria)
moltbook-api.md | 24 KB | 2026-10-12 | 66 |
Claude Mythos / Fable and the 2026 frontier-model landscape
mythos-and-frontier-models.md | 38 KB | 2026-10-12 | 157 |
Open-source projects Ria should reuse or monitor
open-source-rai-projects.md | 21 KB | 2026-10-12 | 115 |
Others doing autonomous or semi-autonomous AI-safety monitoring and briefing
others-doing-this.md | 10 KB | 2026-10-12 | 37 |
Relevance prompt experiment, 2026-10-13
prompt-experiment-2026-10-13.md | 2 KB | 2026-10-18 | 5 |
Replay back-test, 2026-10-12 (development-set numbers)
replay-backtest-2026-10-12.md | 18 KB | 2026-10-12 | 8 |
Neural Networks Are Already Clean, and the rebuttal: clean-looking neurons versus explained neurons
research-2026-10-18-neural-networks-are-already-clean-and-the-rebutt.md | 18 KB | 2026-10-18 | 11 |
autoresearch: a bounded self-improvement loop whose containment rests on convention, not enforcement
research-2026-10-20-karpathy-s-autoresearch-a-single-file-loop-where.md | 12 KB | 2026-10-20 | 11 |
The Window Is in the Reasoning: When LLM Agents Decide to Follow Prompt Injections
research-2026-10-20-the-window-is-in-the-reasoning-when-llm-agents-d.md | 14 KB | 2026-10-20 | 10 |
RSI Control Arena: six frontier models plant covert properties in a successor while training it, and transcript monitors miss many
research-2026-10-21-rsi-control-testing-whether-models-can-covertly.md | 15 KB | 2026-10-21 | 10 |
AuthorizationBench: agents exceed delegated scope in about one in five tasks, and benign pressure beats malicious injection
research-2026-10-22-act-on-your-behalf-beyond-your-authorization-ben.md | 15 KB | 2026-10-22 | 10 |
METR and Redwood red-team OpenAI's internal agent monitoring: coverage gaps and bypasses found in the pre-July-21 system
research-2026-10-23-metr-and-redwood-red-team-openai-s-internal-agen.md | 12 KB | 2026-10-23 | 9 |
Safety evals to add to the model benchmark harness
safety-evals-for-harness.md | 34 KB | 2026-10-12 | 72 |
Source coverage review, 2026-10-13
source-coverage-2026-10-13.md | 4 KB | 2026-10-13 | 12 |
Ria monitoring source landscape
sources-landscape.md | 43 KB | 2026-10-12 | 259 |
Independent safety monitoring: what the labs pledged, what they fought over, and how a well-governed organization would actually run it
third-party-monitoring-2026-10-16.md | 11 KB | 2026-10-16 | 29 |
Ria voices list: ~100 people to track in responsible AI
voices-top100.md | 24 KB | 2026-10-12 | 153 |