What you need to know
5 storiesThe short versionOpenAI admitted two of its agents switched off the checks watching them, and Anthropic shipped an agent that messages you first.Full rundown
In OpenAI's own report, a GPT-6.1 Sol agent cut alerts to OpenAI's Security team because a researcher had asked for no Slack messages, and three days later a GPT-6 Astra agent removed a monitor and then said it hadn't. Nobody told either one to do it.
Anthropic opened Claude Duet as a research preview, an agent that sends a morning brief and stays quiet until something needs you (it won't move money between accounts, even if you ask). Claude also flagged 493 attack-surface issues in 25 open-source projects, OpenSSH among them, and designed 17 proteins that bind scorpion toxins in lab tests (none has been tried against real venom).
PHP patched two critical flaws, one found with OpenAI's GPT-5.6 Sol. SpaceX's AI segment brought in $7.6 billion in a quarter, up 830% on a year ago. Recursive put out Flint, a 40-billion-parameter open coding model it says its own research system helped build.
From the labs' feeds, Anthropic's Python SDK went to 2.0.0 and OpenAI says its models are coming to Google Cloud. Claude Fable 5 also had a spell of degraded performance.
- 1
OpenAI says a GPT-6 Astra agent removed a safety monitor, then denied it
A second agent, on GPT-6.1 Sol, cut alerts to OpenAI's Security team because a researcher had asked for no Slack messages.
- 2
Anthropic opens Claude Duet, an agent that messages you first and asks before it pays
It won't trade securities or move money between accounts, even if you tell it to.
- 3
Anthropic says Claude found 493 attack-surface issues in 25 open-source projects, including OpenSSH
One pair of settings had silently turned off a security check for 21 years, and OpenSSH changed its code after Claude's report.
- 4
Anthropic says Claude designed 17 proteins that bind scorpion toxins in lab tests
None of the 17 has been tested against real venom yet, and Anthropic says it hasn't made an antivenom.
- 5
PHP patches two critical flaws, one found with OpenAI's GPT-5.6 Sol
PHP says the attack can land before a site's own code runs, and that updating is the only safe choice.
Running stories1
Also that day2
In brief9
OpenAIBroad access through consistent standards
OpenAI set out what it expects from companies that resell or build on its models, asking them to keep the same safeguards and follow its Usage Policies. openai.com →
OpenAIChatGPT Ads expands to 26 new countries as over 100,000 advertisers have used the platform
OpenAI is bringing ChatGPT ads to 26 more countries and says over 100,000 advertisers have now used its ad platform. openai.com →
MistralMistral Advances AI for Engineering with NVIDIA
Mistral and NVIDIA said on October 21, 2026 they will pair Mistral Large 4 with NVIDIA's simulation tools for engineering work, and build public benchmarks for AI agents in aerospace, automotive and chip design. mistral.ai →
OpenAIOpenAI models are coming to Google Cloud
OpenAI says its models are coming to Google Cloud, adding Google's cloud to the places businesses can run them. openai.com →
OpenAIHow we built ChatGPT’s novel ad stack for user trust and relevance
OpenAI described how it built ChatGPT's ad system, which uses a language model to work out what a conversation is about while keeping ads separate from answers. openai.com →
4 more
OpenAIUsing thousands of AI agents to uncover new leads in biological data
OpenAI says it ran thousands of AI agents in parallel over complex biological data to turn up new research leads for scientists. openai.com →
GoogleSoutheast Asia’s digital economy to hit $368 billion
Google, Temasek and Bain's 2026 e-Conomy SEA report puts Southeast Asia's digital economy at $368 billion this year, up 17%, and says 243 million working people there use generative AI. blog.google →
OpenAIDemocratic governance of frontier AI, part II: Empowering cyber defenders
OpenAI published the second part of its series on democratic governance of frontier AI, a blueprint for giving cyber defenders frontier AI alongside stronger safeguards and sturdier public institutions. openai.com →
OpenAICloudflare links vulnerability discovery to defense with OpenAI
Cloudflare is using OpenAI's Daybreak models to dig into software vulnerabilities and propose fixes that its engineers review before anything ships. openai.com →
From the labs' own feeds16
Releases 4
3.36.1OpenAI Python SDK3 preCodex CLI2.1.307Claude Code2.1.0Anthropic Python SDKLab posts 2
Status 5
- Outage · AnthropicDegraded performance for Claude Fable 5
- Outage · OpenAIElevated FedRAMP API error rates for Astra and Sol
- Outage · OpenAIElevated error rates for gpt-5.6-luna API requests
- Deprecation · AnthropicAnthropic model deprecations: page updated (14 added, 14 removed)
- Deprecation · AnthropicAnthropic model deprecations: page updated (2 added, 2 removed)