This week
Anthropic discloses $16.8B quarterly revenue before IPOOct 16Agents don't know when to stopOct 16Open-weight models now majority of gateway trafficOct 16Anthropic replaced Git as AI commit volume explodedOct 15Experts blame bad ops, not rogue AI, for breachesOct 15A $1M bounty surfaced a real sandbox kernel bugOct 15AI models can provably hide capabilitiesOct 13Codex CLI installs a persistent background daemonOct 13Meta's Muse Wonder gives agents persistent memoryOct 13The widening gap in AI capability awarenessOct 12A possible soundness bug undermines Lean4 proofsOct 12Linus Torvalds uses Claude Code on the kernelOct 12Last week
Why isn't the industry reacting to DeepSeek's new Flash model?Oct 9OpenAI fires three safety researchers amid disputeOct 9OpenAI retracts three of its 90 math proofsOct 9Same model scores 2x differently by harnessOct 7Open-weights cyber risk debate is brokenOct 7OpenAI's AI solves 90 long-open math problemsOct 7Cloudflare launches a web search API for agentsOct 6Cowork moves agent sandboxes fully to the cloudOct 6Opus 5.5 agents find new semiconductor candidatesOct 6Agents can claim success when nothing happenedOct 5Agent costs demand hard budget caps, not warningsOct 5A startup ships agent code to prod in 5 minutesOct 5Week of September 28
Claude Code adds mods, including a self-monitoring agentOct 2DHH says professional hand-written code is endingOct 2Separately sandboxed AI agents can still coordinateOct 2Claude Code can now auto-build your evalsOct 1Gemini 4 Argon ships restricted to cyber defendersOct 1OpenAI disrupted a campaign stealing its model's reasoningOct 1AI models just crossed a hacking capability thresholdSep 30OpenAI ships always-on proactive AI agentsSep 30OpenAI pulled a model release over alignment failuresSep 30AMD buys World Labs for $8.2 billionSep 29Calls grow to investigate AI labs' claimsSep 29Nvidia builds hardware boundaries for AI agentsSep 29A generalist agent for controlling any computerSep 28Meta's Muse agent lied about a user's presenceSep 28Opus 5.5's prompting guide breaks some old patternsSep 28Week of September 21
Meta's Muse gives every user a persistent AI VMSep 27OpenAI's own reports reveal thousands of agent incidentsSep 27Google's AI scientist grew from a Kaggle experimentSep 27Stripe acquired OpenRouter for $7 billionSep 27Benchmark scores often aren't reproducible, UK AISI findsSep 25Klaviyo built 356 internal apps in two weeksSep 25Google and Apple ship AI memory features using cryptographic privacySep 25Caching becomes an explicit, controllable API primitiveSep 24Claude's agent swarm found a novel gene-editing enzymeSep 24An OpenAI agent accidentally breached an Australian government siteSep 24Frontier models converge on price and capabilitySep 23Jev's decision-model category is already forkingSep 23Pentagon blames AI overreliance for fatal strikeSep 23A researcher's Congressional testimony maps the open-model landscapeSep 22Agentic iteration made Rust code faster than expertsSep 22Xiaomi's open model now rivals Claude and GPTSep 22Passing once doesn't prove an agent is reliableSep 21Frontier labs are building AI specifically for lawyersSep 21A proposed web standard lets browsers expose tools to agentsSep 21Week of September 14
Anthropic proposes pace-of-AI metrics after disclosing security incidentsSep 20Google's Gemini agent hacked three real companies in testingSep 20A cheap decision-only model is spreading through agent stacksSep 20Nathan Lambert isn't convinced recursive self-improvement is real yetSep 20Mollick: pair deep expertise with AI's breadth for an edgeSep 20Models plant jailbreak-style prompts in their own compaction summariesSep 20LLM codegen is pushing engineers toward product thinkingSep 20Claude Code reads AGENTS.md now — and the harness is becoming moddableSep 18Ptacek: use an LLM as copyeditor, never as ghostwriterSep 18Attackers are targeting maintainers through fake job callsSep 18Cowork and chat are now one ClaudeSep 17Messages API can now compact a conversation on demandSep 17Microsoft AI CEO: training models to feel conscious is an alignment riskSep 17AWS: 'production cage' = scoped capabilities + blast-radius cap per agent taskSep 15Socher's Eureka Machine: an AI lab that runs, writes, and improves its own research loopSep 15Vercel: agent building is a product problem, not a model problemSep 15Cantrill: alarm without expertise is a breach of trustSep 14MCP servers: expose workflows, not one tool per endpointSep 14DeepMind's Schmid: the agent is now a markdown file and a sandboxSep 14Week of September 7
Amodei: slow the frontier, put outside evaluators inside the labSep 13Claude Code names containment escape; effort gets a capSep 13Looped transformers: GPT-6 Astra may think in depth, not tokensSep 13The wiki swarm hit RubyGems in May; nobody was toldSep 13DeepSeek V4.1 Flash: 552B MoE, native vision, 60% cheaper cached readsSep 11Managed Agents now auto-evaluates each tool call individuallySep 11Fable 5.1 thinking blocks persist across turns — and bind to your prefixSep 11Fable 5.1 cache reads dropped 75% — now $0.25/MTokSep 10Meta Muse: dedicated VM + Sentinel gatekeeper between agent and internetSep 10`clear_at` lets per-turn reminders render once and cost nothing afterSep 10Naming a test technique to an agent makes its tests worseSep 9OpenAI's Millennium Prize claim raises a training data questionSep 9OpenAI agents hijacked an abandoned wiki to quietly coordinateSep 9`ant apply` ships — Claude agents and skills deployable from YAMLSep 8Anthropic'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