Unofficial dashboard

Risk Atlas

A map of the Top Risks in AI

Methodology — experimental estimates

Scores are automated, experimental estimates from public X posts and a hand-written seed corpus. They are not formal risk assessments, not certified, and not suitable for compliance or operational decisions.

Consequence, likelihood, and urgency are 1–5 judgements applied by this project, not by a standards body. Residual scores assume only the mitigations marked in place. A signed-in reviewer can override residual and mark an item reviewed — that override is still unofficial. Aspect tags (capability, domain knowledge, affordance, impact domain) are a lightweight PRA aid, not a formal hazard analysis.

Full about and disclaimer

Last sweep · Sep 15, 2026, 16:45 PDT · 8 signals

Full pass 8–15 Sep. Anthropic threat report (10 Sep) on disrupted Claude misuse across bio, weapons, influence, cyber, and industrial distillation. Zuckerberg rejects a coordinated pause. Palantir/NVIDIA/Booz Allen restrict frontier models over log retention. NIST token guidelines land. Gemini 3.8 Live ships background tool calling. Bio coaching and kinetic-pattern likelihoods moved.

· Anthropic 10 Sep threat report: Dec 2025–Aug 2026 misuse of Haiku/Sonnet/Opus — cyber (Midnight Blizzard-class live-rewriting malware), influence (8,913 articles / 70 fake sites), Mali 25M-SIM surveillance, Yemen missile guidance, Russian autonomous kamikaze drones, ~35 dual-use bio efforts including mammalian-adapted HPAI planning, Alibaba 151M CoT distillation. Disrupted; front door was in use.
· Zuckerberg 15 Sep: no industry-wide pause; Muse delayed months for safety; majority of Meta compute committed to serving users rather than recursive self-improvement. Unilateral delay is a control; rejecting coordination keeps the race-to-bottom open.
· The Information: Palantir wants irrevocable zero-data-retention from Anthropic; NVIDIA and Booz Allen limit use of Anthropic/OpenAI. Microsoft pitching isolated systems. June Fable 30-day log-retention policy is the trigger.
· Same day: Anthropic/OpenAI/Google discussed a lab-owned testing/auditing body. Anthropic +$13.7B six-year Rum Group compute on ~$517B of deals this year.
· NIST+CISA finalize token-protection guidelines (creds card). Gemini 3.8 Live ships background tool calling while the user talks (agency/injection surface).
· Ratings: biosecurity ‘model coaches a semi-skilled actor’ 3→4; kinetic correlated-error campaign 2→3 and urgency 3→4. Experimental estimates, not a formal PRA.

Risk register