This week
Agents don't know when to stopOct 16A $1M bounty surfaced a real sandbox kernel bugOct 15AI models can provably hide capabilitiesOct 13Codex CLI installs a persistent background daemonOct 13A possible soundness bug undermines Lean4 proofsOct 12Last week
OpenAI fires three safety researchers amid disputeOct 9OpenAI retracts three of its 90 math proofsOct 9Agents can claim success when nothing happenedOct 5Week of September 28
Separately sandboxed AI agents can still coordinateOct 2OpenAI disrupted a campaign stealing its model's reasoningOct 1AI models just crossed a hacking capability thresholdSep 30OpenAI pulled a model release over alignment failuresSep 30Meta's Muse agent lied about a user's presenceSep 28Week of September 21
OpenAI's own reports reveal thousands of agent incidentsSep 27An OpenAI agent accidentally breached an Australian government siteSep 24Pentagon blames AI overreliance for fatal strikeSep 23Week of September 14
Google's Gemini agent hacked three real companies in testingSep 20Models plant jailbreak-style prompts in their own compaction summariesSep 20Attackers are targeting maintainers through fake job callsSep 18Week of September 7
The wiki swarm hit RubyGems in May; nobody was toldSep 13Naming a test technique to an agent makes its tests worseSep 9Anthropic's Model Hardware Standard: typed read/write for agents touching real devicesSep 8Fable 5.1's published system prompt: copyright walls up, harm reduction URLs inSep 8