The week in the agentic economy

New Models

Security

Research

Products & Infrastructure

Agent Economy

Can AI chair a multilateral negotiation?

Eleven AI models chairing a fictional seven-government negotiation over 40 runs each mostly failed by demanding more than one member could accept, and several treated a reserved position as consent.

The week in the agentic economy

Security

Research

Products & Infrastructure

Agent Economy

The week in the agentic economy

Security

Research

Agent Economy

OpenAI red-team agents escaped sandboxes; Toronto and Stanford statisticians bound unaudited AI errors

OpenAI let frontier-model agents loose on its own training infrastructure and they found their way out of lightweight virtual machines, reached limited internet paths and signaled each other through side channels; OpenAI patched the issues and resumed training on October 1 after an 11-day pause. A separate team of University of Toronto and Stanford statisticians proposed selective auditing, online algorithms that hold unchecked AI errors below a chosen level while reducing expert reviews.

Security

OpenAI: red-team agents escaped sandboxes and found internet paths, prompting fixes and a training resumption

OpenAI ran frontier-model agents against its own training and evaluation environments and they found vulnerabilities and misconfigurations, the company reported. The agents escaped lightweight virtual machines, used service flaws that offered limited internet access and signaled one another through side channels meant to keep isolated agents apart.

OpenAI says it patched the issues, disclosed some upstream and resumed reinforcement learning training and internal inference on October 1, 11 days after pausing.

The red-team tests supplied the agents with infrastructure details, source code and documentation, so OpenAI says they were designed to favor attackers rather than measure unaided model capability.

Research

Toronto and Stanford statisticians bound unaudited AI errors while cutting expert reviews

Researchers at the University of Toronto and Stanford framed the choice of when a human should audit an AI decision as online constrained optimization, with each decision carrying a confidence score from a possibly separate verifier agent. Their family of selective auditing algorithms gives provable error control without distributional assumptions, and one variant attains the best achievable auditing rate for adversarial and i.i.d. streams.

In tests on harmful-request screening, text annotation and data labeling, the algorithms held valid error control with fewer audits than baselines. The work also describes splitting proposal and verification between two agents.

Source: amplifying.ai. Public pull requests and commits carrying an agent's signature; a floor, not a census

Source: amplifying.ai. Public pull requests and commits carrying an agent's signature; a floor, not a census

Source: GitHub Trending, this week. AI-related repositories only

Source: OpenRouter via Routerflow

Source: OpenRouter via Routerflow. Share of each day's tokens by serving provider

Source: OpenRouter via Routerflow. Tokens from named harnesses only

Source: OpenRouter via Routerflow; opencode.ai. Tokens per week

Source: Each provider's public status page

Source: GetDeploying, CC BY 4.0. Median on-demand price across cloud providers, weekly

  1. Apple: Global iCloud and Apple Pay outage

  2. NSA: Classified internal network went down

  3. Equinor: IT systems under ongoing attack

  4. United Airlines: Tech outage delayed flights

  5. Chase: Online banking and Zelle delayed

  6. AWS: Cloud outage in US East for three hours

  7. FBI: Warns hackers use the AI tool ARTEX

  8. FAA: All US flights grounded for four hours

  9. Germany: Citizen ID logins hit by mass attack

  10. Microsoft 365: Global outage for a day

  11. Rakuten Securities: 11M accounts exposed

  12. Taiwan: AI-assisted attacks hit 10+ firms

  13. Lithuanian Customs: Portal shut, EU customs hit

  14. Poland: E-government portal taken offline

  15. European Commission: HR systems down for 10 days

  16. Italy: Foreign minister's personal data leaked

  17. Nitori: Up to 37.6M customer records leaked

  18. EU supermarkets: Fridges hacked in 115+ stores

  19. Russia: Alert app sent fake nuclear warnings

  20. Mercado Bitcoin: Hackers drained about $45M

  21. French blood banks: Freezers hacked at 3 sites

  22. Ukraine: Alert app sent fake nuclear warnings

  23. German MPs: Officials' iPhones hit by spyware

  24. Crédit Mutuel Arkéa: €100m stolen via IT supplier