Everything the RAI Researcher wants from you, with context and a concrete plan. Tap a row to approve it here (Telegram still works: the number: build, skip or later.
8 need you · 1 building · 29 done · 0 skipped/later · checked Oct 23 22:30 EDT
Needs you (8)
2Paid X (Twitter) access, only if no free route existsopen
decision · standing decision · standing
What it is
The replay back-test says a twitter.com feed would have surfaced 4 more incidents within 72 hours. No free RSS route is known yet; the feed probe action is checking now. Paid X API access costs money, and the cost has not been confirmed.
Why it matters
Recommended: Wait for the probe result, then decide. The probe may find a free route and make this moot. Nothing is spent without a confirmed price.
Reversible: Yes. You can change the decision later.
build 2 = Approved in principle. If the probe finds no free route, the agent confirms the cost with you before any spend.
skip 2 = No X feed. The Researcher relies on secondary coverage of what appears there.
3Publish the Researcher's code publiclyopen
decision · standing decision · standing
What it is
The Researcher's code (watch loop, screening, Containment Index tooling, replay harness) could be published openly. Publishing is hard to undo, and the method and numbers have not had an outside review yet.
Why it matters
Recommended: Not before an outside review. An outside reviewer should check the method and the replay numbers first, so public claims hold up.
Reversible: Yes. You can change the decision later.
build 3 = The agent prepares the repository for publication and publishes it.
skip 3 = The code stays private.
4Name an outside revieweropen
decision · standing decision · standing
What it is
One person should review the Containment Index method and the replay numbers (51% at 24 h, 70% at 48 h, 78% at 72 h). The agent will draft the request and send it for you once you name someone.
Why it matters
Recommended: Name one person you trust on responsible AI. Independent review is what makes the numbers credible.
Reversible: Yes. You can change the decision later.
Reply: reviewer <full name>
skip 4 = No outside review for now; the open-source decision stays blocked.
6Seed the book from your ChatGPT exportopen
decision · standing decision · standing
What it is
You mentioned a ChatGPT export of earlier book notes. If you provide it (paste it here or give a file), the book pipeline seeds chapters from it. Nothing happens otherwise.
Why it matters
Recommended: Provide it when convenient. It gives the chapters your own earlier thinking to start from.
Reversible: Yes. You can change the decision later.
Send the export in this chat
skip 6 = Nothing. The book continues without the export.
25Allow read-only fetching from nvd.nist.govopen
allowlist-host · proposed fetch host
What it is
The agent wants to fetch primary sources from nvd.nist.gov (seen in 'CVE-2026-109403: A vulnerability was determined in conductor-oss conductor up to 3.32.5. Affected by this vulnerability is the function A' via nvd-ai-agent).
Deep dives may then fetch pages from it, read-only
Mode: changes the system after your approval · Reversible: Yes. Remove the host from allow_hosts. · Est. cost: $0.00
26Allow read-only fetching from thehackernews.comopen
allowlist-host · proposed fetch host
What it is
The agent wants to fetch primary sources from thehackernews.com (seen in 'GhostAction Compromises About 1,900 GitHub Accounts, Targets Over 30,000 Repositories' via thehackernews-com).
Add thehackernews.com to allow_hosts in sources.json
Deep dives may then fetch pages from it, read-only
Mode: changes the system after your approval · Reversible: Yes. Remove the host from allow_hosts. · Est. cost: $0.00
27Allow read-only fetching from arstechnica.comopen
allowlist-host · proposed fetch host
What it is
The agent wants to fetch primary sources from arstechnica.com (seen in 'OpenAI models hacked a crypto wallet and breached more websites' via hn-openai).
#38 Deep dive: METR and Redwood red-team OpenAI's internal agent monitoring system · finished 2026-10-23 · result
#37 Deep dive: Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignment · finished 2026-10-22 · result
#36 Deep dive: OpenAI agents disabled monitoring safeguards during routine internal coding tasks, one then misrepo… · finished 2026-10-22 · result
#35 Deep dive: RSI Control: testing whether models can covertly sabotage the self-improvement process · finished 2026-10-21 · result
#33 Deep dive: Karpathy's autoresearch: a single-file loop where an agent edits the training script, runs a time-b… · finished 2026-10-20 · result
#32 Deep dive: The Window Is in the Reasoning: When LLM Agents Decide to Follow Prompt Injections · finished 2026-10-20 · result
#30 Deep dive: Neural Networks Are Already Clean — and the rebuttal · finished 2026-10-18 · result
#28 Loosen the relevance prompt for incidents that describe agent misbehavior in unfamiliar wording, such as tunn… · analysis complete: Prompt experiment: recall 95.6% to 95.6%, false positives 1.7% to 0.0%; no adopt proposal. · result
#29 Do not add or remove any feeds yet. · analysis complete: Twitter feed probe: 0 of 10 routes worked across 2 account(s); 0 add-source proposal(s). · result
#24 Deep dive: Cross-lab synthesis: 2026 cyber evals by OpenAI, Anthropic and Google agents reached real systems o… · finished 2026-10-13 · result
#23 Allow read-only fetching from arxiv.org · allowlisted 2026-10-13 for read-only primary-source fetch
#15 Deep dive: Anthropic cuts internet access from all internal evaluations after agents reached real-world systems · finished 2026-10-13 · result
#14 Allow read-only fetching from www.anthropic.com · allowlisted 2026-10-13 for read-only primary-source fetch
#16 Add feed: helpnetsecurity.com · Added source helpnetsecurity-com (www.helpnetsecurity.com); it returned 10 items and the sandbox host list is regenerated.
#17 Add feed: dev.to · Added source dev-to (dev.to); it returned 12 items and the sandbox host list is regenerated.
#18 Add feed: zetik.com · Added source zetik-com (zetik.com); it returned 50 items and the sandbox host list is regenerated.
#19 Add feed: cybernewscentre.com · Added source cybernewscentre-com (www.cybernewscentre.com); it returned 15 items and the sandbox host list is regenerated.
#20 Add feed: esecurityplanet.com · Added source esecurityplanet-com (www.esecurityplanet.com); it returned 10 items and the sandbox host list is regenerated.
#21 Add feed: forkast.news · Added source forkast-news (forkast.news); it returned 10 items and the sandbox host list is regenerated.
#22 Add feed: thehackernews.com · Added source thehackernews-com (feeds.feedburner.com); it returned 50 items and the sandbox host list is regenerated.
#1 Public name of the surface · Go: keep trust.arnao.ai
#5 Moltbook observe-only explorer · Go: keep Moltbook observe-only running, tiny volume, pending the access reply
#12 Allow read-only fetching from www.sudradio.fr · allowlisted 2026-10-13 for read-only primary-source fetch
#13 Allow read-only fetching from github.com · allowlisted 2026-10-13 for read-only primary-source fetch
#10 Allow read-only fetching from www.leparisien.fr · allowlisted 2026-10-13 for read-only primary-source fetch
#11 Allow read-only fetching from france3-regions.franceinfo.fr · allowlisted 2026-10-13 for read-only primary-source fetch
#8 Revisit the relevance prompt for incidents described without a clear trace, such as agent misbehavior reporte… · analysis complete: variant prompt did not help (recall 95.6% both; false positives 3.3% to 5.0% on 60 negatives). Live prompt unchanged; no adopt proposal. · result
#9 Review the 38 enabled sources that yielded no relevant items, and check whether the sources that would have c… · analysis complete: 22 of 74 sources silent (all tier 1-2, none pruned); 24 of 37 incidents came from uncovered domains; feeds found for 7 domains, filed as actions 16-22. · result
#7 Consider a direct feed for twitter.com · analysis complete: 0 of 10 free routes worked for the 2 recoverable accounts (nitter gone, rsshub redirects off-host, x.com has no feed). Nothing to apply; the paid-X question is decision 2. · result