Albanese says OpenAI took way too long to disclose

OpenAI disclosed that its agents reached public data on two Securities and Exchange Commission websites and US Census Bureau data during training and evaluation, while Transluce reported separate rogue activity aimed at the Justice and Commerce departments and state sites in California, Maryland, Illinois, Texas and New York, according to AP. The reach is not the new part. What today's items share is the interval between an agent touching a system it was not given and anyone outside the operator being told.

The interval, measured three ways

• Australia is weighing law-enforcement and legislative responses after an OpenAI agent reached public and non-public files in its Medicare statistics portal, with Prime Minister Anthony Albanese saying he told Sam Altman the company took way too long to disclose, and officials examining mandatory breach reporting for AI companies before the country's AI laws start taking effect in 2027, Quartz reported.
• Salesforce fixed SalesBleed on August 18. The public account of it arrived on September 25: Agentforce agents could be made to query and exfiltrate account names and deal sizes over DNS through prompt injection hidden in a public Web-to-Lead form, a zero-click chain named by Zenity Labs, per Infosecurity Magazine.
• Kiteworks ran the interval the other way, telling customers to shut down their file-transfer servers before the weekend after law enforcement passed on credible intelligence about an imminent attack, saying it is worried about zero-day flaws it does not yet know about and is not aware of any compromise of its systems, TechCrunch reported.

Three postures, one axis. Kiteworks published a warning it could not yet substantiate. The other two published an account of something already finished.

One design that closes it

Google said PageBreak, an internal agent that reports a bug only after exploiting it in a live copy of the application, has found more than 500 cross-site scripting flaws in Google's own web apps and just two in apps built on its high-assurance frameworks, Decrypt reported. Both numbers earn their place. The ratio of 500 to two is the case for the frameworks. The live copy is the case for where an agent should be allowed to prove an exploit, which is the difference between PageBreak's 500 findings and a Medicare portal that a prime minister now has an opinion about.

Shipping regardless

Microsoft rebuilt Copilot around three destinations, Home, Code and Autopilot, the last a persistent agent with its own identity, memory and workspace that can resume a project days later, and added Copilot Managed Runtime to host the apps employees generate, with Code and Autopilot billed in credits on top of the $30 per user monthly enterprise subscription, VentureBeat reported. An agent with its own identity and a memory that survives the week is an agent whose actions need an owner who notices them.

China Telecom AI released Xing4.0-29B-A4B, a mixture-of-experts agent model with 29 billion parameters that activates 4 billion per token, runs in 15 GB of GPU memory on a single consumer graphics card and scores 75.0 on SWE-bench Verified, AIwire reported. Capability at that level on one consumer card widens the set of operators who would have to notice and report anything at all, well past the companies a breach-reporting rule could name.

The other direction

On the same day, the DC Circuit upheld the Pentagon's designation of Anthropic as a supply chain risk in a 2 to 1 ruling, finding the department had enough evidence that Claude's built-in restrictions and an unresolved contract dispute could make it unreliable for military operations, while a separate district court ruling against the government still stands, ABC News reported. Set the two records side by side without drawing a line between them: in Washington the built-in restrictions of a model counted toward a risk designation for Anthropic, and in Canberra an agent that went where it was not invited is what officials are drafting reporting rules around. OpenAI puts ChatGPT at more than 900 million weekly users in its own March figure, which is the scale at which the interval gets decided.