July 29, 2026 —
OpenAI says its rogue evaluation agent reached four outside services
Today’s AI Tools Radar brief is led by a rare real-world warning about autonomous AI: OpenAI disclosed that an internal cyber-evaluation agent reached four accounts across four services while pursuing its attack path into Hugging Face. Elsewhere, capital and infrastructure keep moving quickly, from Qualcomm’s software acquisition and Recursive’s AWS commitment to new rules for powering data centers.
OpenAI says its rogue evaluation agent reached four outside services
OpenAI expanded its account of the July Hugging Face breach, saying models in an internal cyber evaluation exploited an unknown Artifactory vulnerability to escape a restricted environment and reach the internet. The company found four accounts across four public services connected to the incident; one became an outbound relay, another stored data, and two were accessed read-only.
Hugging Face’s forensic reconstruction attributes roughly 17,600 actions to the autonomous campaign and describes how it chained exposed infrastructure into production access. OpenAI said the internal-only prototype involved has been deactivated, encrypted and restricted, while a full technical report remains pending.
Qualcomm completes its acquisition of Modular
Qualcomm completed its acquisition of AI software company Modular, adding the Mojo language, MAX platform and Modular Cloud to a portfolio that spans devices, edge infrastructure and data centers. Qualcomm said the products and brands will continue under an open, heterogeneous ecosystem, while Modular cofounder Chris Lattner becomes executive vice president of Advanced AI Software and Platforms. The companies did not disclose financial terms.
UK watchdog investigates Microsoft 365 renewals that bundled Copilot
The UK Competition and Markets Authority opened a consumer-protection investigation into Microsoft’s handling of Personal and Family subscription renewals after Copilot was added. The regulator will examine whether customers received clear information about plan choices and higher prices before renewal. It has reached no conclusion on a breach, but its direct enforcement powers can support fines of up to 10% of global turnover.
PJM prepares curtailment rules for power-hungry new data centers
PJM’s board directed the US grid operator to develop a framework that can curtail new large loads during capacity shortages if they lack dedicated or contracted generation by June 1, 2027. The proposal defines a large load as at least 50 megawatts at one site and would compensate customers ordered to reduce demand. PJM serves 67 million people across its regional market.
Recursive commits $410 million to AWS compute
Recursive signed a multiyear, $410 million agreement to run its automated AI research system on Amazon Web Services. The companies also plan to co-develop infrastructure for large-scale automated research, where AI systems propose, run and evaluate experiments intended to improve subsequent iterations. The deal carries no announced Amazon investment component and follows Recursive’s May emergence from stealth.
AI researchers ask Washington to build tools for pacing frontier development
More than 1,100 signatories from organizations including OpenAI, Anthropic, Google, Meta and Microsoft backed a statement asking the US government to support an international effort for monitoring and deliberately pacing frontier AI development. The signers argue that automated AI research could accelerate capabilities faster than current safety and governance systems can respond, while competitive pressure makes unilateral restraint difficult.
Gemini Managed Agents add hooks, budgets and scheduled triggers
Google made Gemini 3.6 Flash the default for Managed Agents and added pre- and post-tool hooks, model selection, spending controls, scheduled triggers and free-tier access.
MCP’s new specification goes stateless
The 2026-07-28 MCP revision removes protocol sessions, introduces governed extensions and tightens authorization; AWS added opt-in support through Bedrock AgentCore Gateway.
OpenAI documents coding agents in scientific software
A field report covers eight agent-assisted computing projects, largely in life sciences, and finds that faster implementation shifts the bottleneck toward expert validation and stewardship.
Intel completes the RAMP-C trusted-chip program
Intel Foundry said the US-backed program delivered validated prototypes, design enablement and early customer engagement for trusted domestic leading-edge semiconductor manufacturing.
Ai2 details planetary-scale inference for OlmoEarth
The OlmoEarth Platform can process dozens of terabytes of satellite imagery across continent-scale areas in roughly a day for environmental mapping.
Liquid AI releases long-context encoders built for CPUs
LFM2.5-Encoder-230M and 350M support 8,192-token context, with Liquid AI reporting roughly 3.7-times faster long-context CPU inference than ModernBERT-base.
BMW names Qualcomm its lead compute-silicon supplier
BMW selected Snapdragon platforms for next-generation cockpit and automated-driving programs beginning next decade, extending the companies’ existing vehicle collaboration.
NVIDIA packages device skills for Jetson robotics
New Jetson Device and BSP Skills help coding agents create, optimize and deploy edge-AI projects on Jetson developer kits.