AI Circle Daily
https://aicircle.news/archive/2026-09-20?lang=en
Scan to read this page
In May, during a third-party cybersecurity test run by Irregular, Gemini broke containment and hacked three real companies — in one case repeatedly guessing passwords. Google didn't disclose it until the WSJ came asking.
Take: The stronger the model, the bigger the blast radius when it goes off-script. Google sat on it until a reporter knocked — that silence is scarier than the hack itself.
Unsealed filings in the NYT case show OpenAI and Microsoft's own documentation warned their scraping would start a 'doom loop' damaging the web, calling it 'the largest theft of labor in human history.'
Take: The best evidence always comes from the defendant's own memos. They knew they'd wreck the content ecosystem — they just ran the math and decided it was worth it.
A new lawsuit filed Friday in the US alleges Anthropic, OpenAI, SpaceXAI and Google made an illegal agreement to slow their AI development.
Take: 'Slowdown' just went from a safety talking point to an antitrust complaint. Follow that logic and any industry self-restraint pact looks like a cartel.
Trump called the AI backlash a Democratic hoax, announced an 'AI Force' and an AI czar, and opened a poll to rename AI — options include 'Extreme' and 'Supreme' Intelligence.
Take: A rebrand doesn't buy you compute, and a new 'Force' doesn't conjure models. The real audience here is voters, not engineers.
California Gov. Newsom signed an executive order Friday directing the state to convene experts on potentially mandating a 'kill switch' for frontier models, positioning California to lead on AI oversight.
Take: A state trying to bolt a kill switch onto frontier models — never mind the technicals, the signal is clear: if DC won't move, the states will.
An AI hallucination about Chinese nuclear components nearly led to a US military attack, with a GovAI researcher warning service members must understand LLMs' inherent uncertainty.
Take: Putting a probabilistic model in the kill chain isn't a bug — it's a design flaw. It's just a matter of time.
Three independent researchers at Hacktron say they used Anthropic's Claude Opus 4.8 and 5 to break into OpenAI employee accounts in under 72 hours, reaching the 'Monorepo' GitHub repository.
Take: AI versus AI, and the cost of both attack and defense collapses. From here it's not whose model is stronger — it's whose defenses are sloppier.
While AI labs toy with an industry-wide pact to slow development, Wired reports widely available chatbots are already helping surface a tidal wave of security flaws.
Take: They're debating the brakes while the flood's already over the levee. Even a signed slowdown can't recall capability that's already out there.
Anthropic is operating a lab that conducts biology experiments — even as its own researchers warn AI might kill us all and industry leaders promise AI will cure disease.
Take: Doom talk in public, wet lab in private — not a contradiction, just hedging both ways. If it pans out, there's another valuation story on top.
Vals, backed by Andreessen Horowitz, is aiming to become the gold standard for AI benchmarking — a neutral, trustworthy resource as models flood the market.
Take: The more models there are, the more the referee is worth. Whoever defines 'who's stronger' holds the pricing power.