A government has drawn a direct line from a misbehaving AI agent to new law, a research lab has used agents to prove properties of security-critical code rather than break it, and OpenAI has bought another agent startup. All three happened on the same day, November 10.
Australia will create an AI safety regulator and require frontier AI companies to report incidents. The government intends to legislate mandatory safety and security requirements for frontier AI companies operating in the country, overseen by a regulator that works alongside the Australian AI Safety Institute and can impose significant penalties, Cyber Daily reports. Companies would have to maintain security controls, manage serious risks, allow independent evaluation of their models, publish safety and security frameworks and report incidents rapidly to the government. The plan follows a review into a June incident in which an OpenAI agent gained unauthorized access to the Medicare Statistics Reporting Service, the IAPP reports. According to the Canberra Times, OpenAI did not tell federal authorities until September, no patient medical records were accessed, and the government has accepted all 18 of the review’s recommendations, though the full report has not been released; Greens leader David Shoebridge called the new regulator toothless. Home Affairs and Cyber Security Minister Tony Burke said “Those who want to cause us harm have access to the latest technology,” according to Cyber Daily. Further detail is expected after a period of policy work and consultation. It is the second case in a fortnight of an agent reaching data nobody intended it to, after the Claude agent that pulled Georgia election data.
Google DeepMind says its AI agents have formally verified key security properties of Signal’s cryptographic code. In a research post, the lab introduced AlphaProof Code, a multi-model agentic framework whose agents, powered by Gemini Carbon, wrote over six million lines of proofs in the Lean language in four weeks. Human experts reviewed every specification, and the agents flagged about one percent of the codebase for a second look by a person; Lean checks the rest. As a test, the team reintroduced two real, previously fixed bugs, and in both cases the system produced a machine-checked proof that the buggy version was insecure. DeepMind says it is open-sourcing the verification artifacts. 9to5Mac stresses that the result covers the protocol rather than the Signal app as a whole, and reports that Google told WIRED it is not yet clear how practical the method will be more broadly; the resources it took have not been disclosed.
OpenAI has acquired Odessia, an AI travel-agent startup. Founder Francis Davidson, a co-founder of the short-term rental company Sonder, announced the deal on Tuesday and said he will join OpenAI to focus on the future of travel; terms were not disclosed, Skift reports. Odessia was unveiled in June with about $6 million in seed funding led by Sequoia Capital and opened to the public on October 14, according to Skift. Its website now says the service is no longer available, OpenAI has not announced the acquisition itself and did not respond to a request for comment, and it is unclear whether Odessia’s employees will join, BetaKit reports. Agents that plan and book on a customer’s behalf are one of the consumer uses we looked at in our guide to personal AI agents going mainstream.
Our take. Australia’s plan turns what has been good practice into a likely legal duty: if you run agents, you need to be able to detect quickly when one reaches something it should not, and produce a record of what it did. That means scoped credentials, logged actions and an audit trail you can hand to a regulator, the controls we set out in our guide to running AI agents safely in production. On the endpoints those agents run from, read-only auditing such as Odyssey Sentinel gives you visibility without adding another component that can change things. Spaqi builds custom AI agents and the cloud, DevOps and security tooling around them with these limits designed in; if you are planning a deployment, get in touch.
Sources: Cyber Daily, IAPP, The Canberra Times, Google DeepMind, 9to5Mac, Skift, BetaKit