Desk
Security, safety, and adversarial ML.
New tests on simulated corporate networks found DeepSeek completing 73% of intrusion tasks against 81% for OpenAI’s restricted Daybreak model. That gap is closing faster than the gated-release strategy assumed.
OpenAI is bringing Hacktron's AI-assisted offensive researchers inside the perimeter — focused on third-party service seams — and expanding their access to frontier models through Daybreak Red.
After midterm voters pushed back on data-center growth, OpenAI says it will not use NDAs with public officials and will pay full power costs under new community commitments.
OpenAI’s specialist security model scores 48.5% on ExploitGym, helped produce exploits for 948 Linux kernel bugs and 81 new Firefox issues, and is rated Critical. It also stayed in scope on honeypot tests 99.93% to 100% of the time. Access breadth is the open question.
AT&T’s Connected Life security cameras can now ask Gemini what happened at home — natural-language video search and daily summaries — but only if you’re on the $29.99 professional monitoring plan.
OpenAI’s cyber pipeline validated hundreds of Linux local privilege escalations, wrote patches, and argued that unpatched low-priv access is already root for a serious attacker — defense now means isolation and faster fixing, not hope.
OpenAI's cyber model is moving into the government cloud, and the useful part is less the model than the gate around it: verified defenders, monitoring, and data that stays put.
Cisco's case for AI on defense is sound, but this is a customer showcase with plenty of conviction and no numbers to check it against.
OpenAI is turning its own models and an outside firm loose on its defenses again, and the approach is sound, though the post is long on intent and short on measurable results.
OpenAI's much bigger bug bounties are a smart, overdue bet on outside researchers, and the bonus structure rewards the careful reports that actually get bugs fixed.
OpenAI hired a former Army secretary to help sell AI to allied militaries and security agencies, a deliberate step deeper into defense work.
Four big labs built AI bug hunters and landed on the same rule: the agent that finds a flaw never gets to confirm it. That rule is the real story.