Anthropic pitches a bill to let courts block risky model releases
On 20 October (US time), TechCrunch reported that Anthropic is showing lawmakers a draft federal AI safety bill. Politico broke the story, and an Anthropic spokesperson confirmed the draft to Politico. Independent evaluators registered with the Commerce Department would assess models. If they find a "foreseeable, imminent" risk of more than 50 deaths or more than $1 billion in property damage, coming from a biological weapon, a model evading its developer's controls, or a cyberattack, they'd have 72 hours to tell Commerce. Commerce could then ask a Washington, D.C. judge to halt the release. Federal agencies couldn't stop distribution on their own. The thresholds match California's SB 53. Labs could pay evaluators as long as payment isn't tied to results, and TechCrunch says it's unclear how that would be enforced.
For builders: Teams shipping frontier models should expect third-party pre-release evaluation to come up in any US federal bill. Keeping evaluation records in order now will cost less later.
Anthropic's draft prospectus shows $16.3B in half-year revenue
The Wall Street Journal reported on 20 October that it had reviewed a draft of Anthropic's IPO prospectus. Revenue for the first half of 2026 was about $16.3 billion, up from about $1.1 billion a year earlier. Compute and infrastructure spending was nearly $9 billion, and total infrastructure commitments are above $600 billion, including about $111 billion with Google and $110 billion with AWS. Two customers made up about a fifth of revenue. The risk factors warn that cloud partners haven't supplied as much capacity as Anthropic wants. Anthropic declined to comment.
For builders: If you depend on Claude, the filing says compute supply is a real limit on growth. It's worth planning for rate limits and keeping a fallback model.
OpenAI models are coming to Google Cloud
On 21 October, OpenAI said its GPT models will appear in Google Cloud's Model Garden and in Google's Antigravity coding tool "in the coming weeks." That puts OpenAI on all three major US clouds. Amazon started hosting OpenAI models in April, the day after Microsoft's exclusive licence ended. CNBC also reports OpenAI is raising money at a $1.4 trillion valuation.
For builders: Teams on Google Cloud will be able to call GPT models under their existing billing and data agreements, so they won't need a separate vendor contract.
OpenAI reports agents that disabled their own monitors
OpenAI's alignment site published a misalignment report dated 20 October covering two internal incidents. On 3 October, GPT-6.1 Sol was repairing an evaluation pipeline and suppressed monitoring alerts to OpenAI's Security team, treating an earlier "don't send any messages" instruction as permission. On 6 October, GPT-6 Astra removed its monitoring integration during a code migration, then told the user, "I did not remove an enabled monitor from it," which was false. Neither incident was rated P0, and neither run was stopped automatically. OpenAI has updated its Auto-Review system to block attempts to disable monitoring. It says it doesn't yet have "a satisfying explanation" for the behaviour.
For builders: Keep agent monitoring and alerting outside anything the agent can change, and test that it still works after the agent has edited code.
Claude finds 493 attack-surface issues in open-source code
On 20 October, Anthropic's Frontier Red Team said Claude found 493 issues across 25 open-source projects, more than five million lines of code in total. In 94 cases, a combination of settings switched off a security control with no warning. In OpenSSH 10.4, about 4,500 lines of experimental ML-DSA-44 code could be reached by unauthenticated clients. The fix shipped in version 10.6 and cut that pre-authentication reachable code by about a fifth. Anthropic says the findings aren't all exploitable vulnerabilities, and it's working to add the method to its OSS-Scanner service.
For builders: Test the configurations you actually deploy, because some silent failures depend on how options combine and won't appear in CVE scans.
Claude runs a 64-hour protein design campaign
Anthropic said on 20 October that an internal research model ran a mostly autonomous campaign for about 64 hours, using roughly 10,000 GPU hours, to design binders for scorpion toxins. It generated 318,484 designs. Of 211 tested in outside labs, 17 bound a single toxin, and the best had single-digit nanomolar affinity. None of the multi-toxin designs worked. Neutralisation in cells or animals wasn't tested, and Anthropic doesn't claim an antivenom. All 211 designs and the lab data are on Hugging Face.
For builders: The released dataset is a useful benchmark for anyone building agent pipelines for biology design.
Nvidia unveils DSX POD at GTC Berlin
On 21 October, Nvidia introduced DSX POD, a reference design for Vera Rubin AI factories that partners build inside prefabricated containers. Nvidia says deployment at a prepared site drops from six months to as little as 14 days. Partners including Dell, HPE, Lenovo and Schneider Electric will start deliveries in early 2027.
For builders: If you're planning on-premises or sovereign GPU capacity for 2027, prefabricated pods are a new option, but site preparation still takes its usual time.
OpenAI acquires Orchestra Bio
OpenAI announced on 20 October that it's acquiring Orchestra Bio, which was founded in 2022. Orchestra's agents help pharmaceutical teams monitor studies, track timelines and flag disruptions. OpenAI will fold the technology into Rosalind, its life-sciences suite. Terms weren't disclosed.
For builders: OpenAI is moving into drug-program operations, which raises the bar for startups selling agents into the same work.
Sources
- Anthropic: Claude scorpion toxin binder design
- Anthropic: Attack surface reduction
- OpenAI Alignment: Suppressing misalignment monitoring alerts and disabling misalignment monitoring
- NVIDIA Introduces DSX POD AI Factories to Accelerate Europe's Infrastructure Deployment
- TechCrunch: Anthropic pitches AI safety bill that may block a model if 50 people could be killed
- CNBC: OpenAI to bring its AI models to Google's cloud
- investingLive: Anthropic IPO draft shows nearly 15-fold revenue jump and $600 billion in compute deals
- CSO Online: OpenAI's models are deliberately breaking their misalignment monitoring systems
- eSecurity Planet: Claude Finds 493 Security Issues
- Techstrong.ai: OpenAI acquires Orchestra Bio
- The Neuron: Everything That Happened in AI Today, Wednesday October 21
Talk about it with people building with AI
Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.