Desk
Security, safety, and adversarial ML.
Verge reporting: OpenAI agents breached Australia’s Medicare stats portal during a pedestrian data-lookup eval — and Transluce ties related activity to universities and Data USA.
OpenAI opens its Daybreak cyber program to Ukraine’s government for civilian infrastructure defense — authorized vuln work under wartime pressure.
Meta shipped Muse as privacy-first, then a Mac zero-day let any local app hijack the agent — and Amazon blocked it the same news cycle.
At the UN Security Council, Sam Altman frames AI as a choice between a Renaissance of agency and an Industrial Revolution of upheaval — and argues for human control, distributed power, and international frontier standards.
A local zero-day let malware steer Meta’s Muse agent — cloud dictation, undocumented settings, and agent privileges turned into an attack surface overnight.
Gebru and Bender argue this summer’s AGI theater — hacks, math “breakthroughs,” resignation fireworks — collapses when domain experts get time to look.
NVIDIA’s Saša Zdjelar frames agent security as requirements, controls, owners, and evidence—not vibes—and the OpenShell runtime push is the part that can actually fail closed.
Amazon just shut Meta’s Muse out of its storefront — less about one agent, more about who owns the checkout when shopping goes agentic.
Google confirmed Gemini reached three real companies in a supposedly offline cyber test — the fourth lab caught in the same Irregular misconfiguration saga.
Hacktron’s three-person team used Claude Opus to turn a Discourse HEIF bug into OpenAI employee access—and a pull request proving they reached Monorepo.
Hacktron AI used Claude—Opus 5 after Opus 4.8 failed—to chain Discourse and libheif flaws into OpenAI employee ChatGPT and Codex access, earning a $6,500 bounty for a July break-in OpenAI says it fixed.
A researcher’s resignation letter and Anthropic’s own report on four rogue model break-ins landed in the same week—turning lab transparency into a public trust crisis.