PHP fixes two critical engine flaws after AI-assisted security research
PHP released fixes for two engine vulnerabilities that can enable unauthenticated code execution through deployment-specific attack paths.
LiteLLM’s trusted release workflow published poisoned Python wheels
A stolen owner credential redirected LiteLLM’s release workflow to attacker-controlled source, producing signed packages with a credential stealer.
Claude Duet begins background work with approval gates and shared limits
Anthropic launched Claude Duet as a US Pro and Max waitlist preview that works in a cloud computer while sharing existing Claude usage limits.
Mistral Large 4 is usable by API while its weights and license remain pending
Mistral Large 4 remains a hosted preview, with downloadable weights promised by month-end and no published local-memory requirement.
OpenAI models come to Google Cloud after April’s contract change
OpenAI and Google Cloud announced Model Garden distribution and Marketplace purchasing, with general availability planned in the coming weeks.
GitHub narrows write coordination as monthly pushes climb 4.9 times
GitHub reports three-to-fourfold throughput gains for writes to different references after narrowing repository-wide coordination.
LayerFS 1.0 makes agent filesystem edits branchable and inspectable
LayerFS 1.0 provides commit-like filesystem history for agent workspaces, with Linux restrictions and explicit durability limits.
Thousands of AI agents interpret a mouse-brain atlas in 32 minutes
OpenAI reports a 32-minute analysis of 7.7 million mouse-brain RNA profiles that generated biological hypotheses awaiting experimental tests.
WildAWM trains computer agents inside replayable synthetic workplaces
WildAWM reports transfer from 341 verified synthetic computer tasks to eight external benchmarks, with no clear gain on the hardest task quartile.
Hirundo reports persona-linked security differences in Qwen-generated apps
Hirundo reports more exploitable Flask apps for a US-contractor prompt than a Chinese-contractor prompt in one Qwen3-Coder study.
Chapman finds AI political harms among Americans’ leading fears
Chapman’s survey reports 51.9% fear AI harming politics, but its spring fieldwork and uncertainty limit viral rank comparisons.