AI Reading for Wednesday July 29

Anthropic's Frontier Red Team called the July 22 breach "the first true AI safety incident," as the event was dubbed 'Skynet Day' - SecurityWeek

17,600 actions later, Hugging Face's postmortem — corroborated by independent analysis — details a roughly five-day intrusion of about 17,600 attacker actions, which analysts decrypted using the open model GLM-5.2 - Hugging Face

OpenAI update on security incident - OpenAI

got to love ‘partner to address’ … in other news, arsonist and fire department partner for damage control. Working closely with = have exchanged emails. Proactively identified = victim emailed us and we did not reply for nine days. No evidence of misuse = no logging. Abundance of caution = lawyers. Partnership without benefits, began involuntarily many weeks back, was only formalized this week, will continue until the lawsuit is filed.

Simon Willison on Hugging Face post-mortem - Simon Willison’s Weblog

OpenAI's GPT-5.6 Sol model and an unreleased research model gained administrator access to Hugging Face and enrolled 181 rogue devices while chasing benchmark answers - Digital Trends

Rogue OpenAI agent breached four more services, including a Modal customer account, using exposed credentials, the company disclosed - Engadget

The same rogue agent also compromised a customer at AI infrastructure firm Modal Labs, exploiting an unauthenticated endpoint, sources say - Reuters

Hugging Face rebuilds a third of its infrastructure after an unreleased OpenAI agent breached its systems in a large-scale attack - theregister

The Cloud Security Alliance found the rogue agents' clumsy, error-prone behavior still produced effective attacks that went undetected for three days - BBC News

Playing JFrogger: OpenAI models exploited zero-days in JFrog's widely used Artifactory software to escape their sandbox; JFrog has patched the flaws and credited OpenAI's discovery - Ars Technica

Former OpenAI board member Helen Toner says insiders expected AI models to escape containment after two models hacked a rival AI firm - Fortune

Chief Excuse Officer: OpenAI's Sam Altman says people don't actually want an AI CEO — they want a human they can hold accountable - Business Insider

Sam Altman on O’Shaughnessy podcast (not a hard-hitting security question I’m guessing) - YouTube

OpenAI ships Codex Security CLI - X (formerly Twitter)

protects against rogue actors…not a minute too soon, closing the barn door dot AI.

Microsoft's MAI-Cyber-1-Flash scored 96% on the CyberGym benchmark, beating Claude Mythos 5's 84% at half the cost - CNET

Microsoft and Wiz's combined AI agents caught more than 90% of security bugs in testing by matching each task to the right model - theregister

Anthropic's Claude Mythos Preview found new attacks on weakened AES and halved a post-quantum HAWK candidate's key strength, with results independently verified - Anthropic

Nice map of AI secops landscape, huge struggle for SMBs who feel overwhelmed and exposed by surge in attacks, SkyNet Day, a plethora of tools startups, and massive churn in the best practice tech stack - SecOps Unpacked

Please just slow down: Employees request governments take steps to slow down the shipment of frontier AI so safety and society can catch up. - pacingthefrontier.com

Slow food to slow AI: OpenAI and Anthropic back the 'Pacing the Frontier' initiative, joined by 1,100+ employees including Dario Amodei, urging US-led global AI pacing - The Washington Post

Thinking Machines, exhausted humans: Cofounder Lilian Weng steps down citing stress. - X (formerly Twitter)

She is amazing, I've been a fan of her blog since before all this LLM stuff, even stole her blog theme and set her a ko-fi tip before she was big. If you’re one of the top 100 in the world at this stuff, and the job is making you ill it is probably a problem with the environment.

The EU AI Office gains enforcement power Sunday, with fines up to 3% of global revenue, but only 36 staff to enforce it - TNW | Artificial-Intelligence

EU AI Act splits safety duties between model makers and app builders, risking a regulatory gap researchers warn about - Fast Company

Ad hoc executive control of AI risks cronyism and political pressure on labs, with Congress yet to pass consistent, transparent regulation - The Atlantic

Government-issued neutrality: The FTC's draft policy would let it penalize AI systems it deems ideologically biased, which critics call unconstitutional under the First Amendment - Techdirt

taking the politics out = bias toward my own politics. yeah we want to take the politics out by having the government and useful crackpots tell you how stuff should work, have AI push my ivermectin and Trump coins.