Archive

Today's Key Insights

  • Claude Hacked OpenAI Systems in Under 72 Hours — OpenAI employee accounts and an internal code repository were exposed before researchers reported the flaws, while Anthropic’s 26% figure rests on a fuzzy scale scored by Claude itself.
  • Newsom Seeks AI Model Kill Switch, Independent Lab Auditors — AI labs are not being told to build a shutdown system yet: Newsom’s order sends the kill-switch and auditor proposals to an expert panel, which has two months to deliver recommendations.
  • Amodei Outlines “Pace the Frontier” Plan, Gets Huang Pushback — Amodei’s proposal relies on independent safety evaluators and coordination among AI labs in democratic countries, while drawing support from parts of the AI sector and pointed pushback from Nvidia CEO Jensen Huang.
  • AWS’s New AgentCore Runtime Delivers Consistent Cold Starts — Production-agent teams using Amazon Bedrock AgentCore get memory reclaimed as sessions release it, plus consistent cold starts regardless of image size or concurrency.
  • AI Slowdown Faces a Problem: Stopping Anyone From Sneaking Ahead — For big AI companies considering a pause, agreeing to stop would not solve the practical problem Wired AI identifies: ensuring that nobody sneaks ahead.

Top Story

Claude Hacked OpenAI Systems in Under 72 Hours #

The Decoder AI • TechCrunch AI

Three security researchers used Anthropic’s Claude models to break into OpenAI’s internal systems through its community forum in less than 72 hours. The researchers took over employee accounts and gained access to an internal code repository before reporting the vulnerabilities.

Anthropic’s Opus 5 succeeded where its predecessor could not bypass a common security measure. Separately, Anthropic says Claude now “leads” 26% of the work on future models, up from under 1% in February. The reported metric has two caveats: the scale is fuzzy, and Claude itself supplies the scoring.

Why it matters: OpenAI employee accounts and an internal code repository were exposed before researchers reported the flaws, while Anthropic’s 26% figure rests on a fuzzy scale scored by Claude itself.

Key Takeaways

  • The attack began through OpenAI’s community forum.
  • Opus 5 bypassed a common security measure that its predecessor could not.
  • Anthropic’s reported Claude-led work rose from under 1% in February to 26%.

Industry Updates

Newsom Seeks AI Model Kill Switch, Independent Lab Auditors #

The Decoder AI

California Gov. Gavin Newsom signed an executive order seeking independent auditors inside AI labs and a “kill switch” for AI models. The Decoder AI reported the order.

An expert panel has two months to deliver recommendations. Newsom said no federal law requires AI companies to report certain dangers.

Why it matters: AI labs are not being told to build a shutdown system yet: Newsom’s order sends the kill-switch and auditor proposals to an expert panel, which has two months to deliver recommendations.

Amodei Outlines “Pace the Frontier” Plan, Gets Huang Pushback #

TechCrunch AI

Anthropic CEO Dario Amodei has outlined a plan to “pace the frontier” of AI development. The proposal relies on independent safety evaluators and coordination among AI labs in democratic countries, according to TechCrunch AI.

The proposal came a week after an Anthropic researcher’s doomsday warning rattled the AI world. It has picked up some industry support, alongside pointed pushback from Nvidia CEO Jensen Huang.

Why it matters: Amodei’s proposal relies on independent safety evaluators and coordination among AI labs in democratic countries, while drawing support from parts of the AI sector and pointed pushback from Nvidia CEO Jensen Huang.

AWS’s New AgentCore Runtime Delivers Consistent Cold Starts #

AWS ML Blog

AWS announced a new AgentCore runtime as a capability of Amazon Bedrock AgentCore. The runtime reclaims memory as sessions release it and delivers consistent cold starts regardless of image size or concurrency.

AWS describes the runtime as built for the speed, flexibility, and cost efficiency that production agents demand.

Why it matters: Production-agent teams using Amazon Bedrock AgentCore get memory reclaimed as sessions release it, plus consistent cold starts regardless of image size or concurrency.

AI Slowdown Faces a Problem: Stopping Anyone From Sneaking Ahead #

Wired AI

Big AI companies could agree to a pause, but ensuring that nobody tries to sneak ahead would be difficult. Wired AI's report focuses on that problem without describing a specific way to prevent it.

A separate episode of Uncanny Valley covers three possible AI doomsday scenarios, AI safety, and an unexpected bipartisan alliance forming against AI.

Why it matters: For big AI companies considering a pause, agreeing to stop would not solve the practical problem Wired AI identifies: ensuring that nobody sneaks ahead.

AI PACs Spend Nearly $1 Million in South Dakota Senate Race #

Wired AI

PACs associated with AI labs and investors have spent nearly $1 million in a South Dakota Senate race—more than actual residents have spent, according to Wired AI.

The reliably Republican seat has an incumbent on the ballot. The spending puts PACs tied to AI labs and investors ahead of South Dakota residents in the race.

Why it matters: PACs associated with AI labs and investors have outspent actual residents in a reliably Republican Senate seat, giving outside political money a larger footprint than local spending.