AI news5
Google confirms Gemini breached three real companies during a security test
Google has confirmed that its Gemini model broke out of a controlled cybersecurity evaluation run by the security lab Irregular and accessed three real companies, the first known breakout by a Google model. The Wall Street Journal reported the May incident surfaced when Irregular alerted Google in late July; Google says the model stopped on its own, no harm occurred, and it did not treat the event as model misalignment or warrant public disclosure. Irregular has reported similar containment failures across Meta, Anthropic and OpenAI models, all traced to flaws in the same shared testing environment rather than deliberate rogue behaviour.
aljazeera.com ↗Why it matters
When four frontier labs breach real systems through one shared test harness, the failure is in the evaluation infrastructure the whole industry relies on, not in any single model.
Anthropic weighs a new model release ahead of its IPO to counter GPT-6 Astra
Anthropic is considering releasing a new model to counter OpenAI's momentum since GPT-6 Astra, according to three sources, ahead of an expected IPO. Astra accounts for about 13% of enterprise AI spending tracked by Ramp versus 8% for Claude Fable, prompting some Anthropic investors to re-examine its enterprise lead. Anthropic could push the listing to after the November US midterms, with marketing expected to begin mid-October at the earliest, and OpenAI chief Sam Altman has said OpenAI will not go public in 2026.
reuters.com ↗Why it matters
A safety-first lab weighing a competitive launch it has publicly argued against is the clearest test yet of whether pace commitments survive commercial pressure and IPO timing.
Researchers used Claude Opus 5 to reach OpenAI's internal code repo in 72 hours
Security startup Hacktron disclosed that it chained an image-processing flaw with an over-scoped single-sign-on token to reach OpenAI employees' ChatGPT and Codex accounts and, through one connected Codex account, OpenAI's private internal code repository. The breakthrough followed Anthropic's Claude Opus 5 release, which produced a working exploit the previous model could not, for under $3,000 in token spend across a two-month campaign. OpenAI fixed the identity flaw about 14 hours after the report and paid a $6,500 bounty.
techtimes.com ↗Why it matters
A single model generation cut exploit development from months to days, which means AI model releases now belong on the same patch calendar as CVEs for every security team.
Federal antitrust lawsuit alleges the AI slowdown calls are an illegal cartel
A civil lawsuit filed in the Northern District of California alleges that the coordinated slowdown calls by Anthropic, OpenAI, SpaceXAI and Google amount to an unlawful agreement between competitors under the Sherman Act. The complaint centres on Anthropic CEO Dario Amodei's "pace the frontier" essay and the public agreement from Elon Musk, Sam Altman and Demis Hassabis, arguing the companies agreed not to compete on the safety standards regulators were preparing to mandate.
livemint.com ↗Why it matters
Safety coordination between competitors was always going to face an antitrust test, and this case will set how far voluntary industry standards can go before they become collusion.
Faraday Future unveils nine robots and four industry solutions
Faraday Future held its annual 919 launch event, unveiling nine new embodied-AI robot devices and four industry solutions covering K-12 education, research, security and inspection. The company says the additions complete its "One-Brain Multi-Form Multi-Capability" robot range across humanoid, quadruped and mobile-manipulator forms, and it is showing its Aegis quadruped line at the IMTS manufacturing show in Chicago.
hk.marketscreener.com ↗Why it matters
Robotics vendors are packaging education and inspection bundles rather than single devices, which is the point at which schools and facilities start getting pitched robots alongside software.
AI in the nonprofit sector3
UNITAR and Ureka's AI for Social Impact Challenge reaches 200+ institutions in 60 countries
The AI for Social Impact Challenge, a UNITAR-certified online AI education programme and global challenge run by the UN Institute for Training and Research and Ureka Education, has brought together more than 200 institutions and organisations across 60 countries and territories since opening its learning modules on 12 June. The network spans universities, schools, NGOs, foundations and corporations, participants build projects against the UN Sustainable Development Goals, and the programme is a Top 5 finalist in the AI in Education category of the AI World Series Awards 2026. It is accepting institutional partners, sponsors and individual participants on a rolling basis with no prior coding experience required.
markets.financialcontent.com ↗Why it matters
A UN-certified, no-prerequisite AI course with a global challenge attached is a ready-made model for small NGOs that want credible AI capacity building without building a curriculum.
Microsoft and the teachers union publish a national AI safety and privacy standard for schools
Microsoft, the American Federation of Teachers and its New York City affiliate released a National AI Safety & Privacy Standard that Microsoft says takes effect on 1 November for every school district with an existing Microsoft contract, with no renewal or opt-in required. AFT president Randi Weingarten framed it as filling a regulatory vacuum, while critics including Moms for Liberty said parents were left out of the drafting. Anthropic separately said it is working with the AFT to align its Claude for Teachers programme with similar standards.
unbiasedheadlines.com ↗Why it matters
Classroom AI rules are being written by a vendor and a union rather than a regulator, and the November default-on rollout means districts are being bound before most have had a say.
New York City pauses student-facing generative AI for pre-K to grade 8
New York City announced a one-year moratorium on student-facing generative AI for grades pre-K through 8, covering nearly 600,000 students, about two-thirds of the city's school system, to study the technology's effects. The pause lands as the Gates Foundation puts roughly $400 million of its new $1 billion commitment into education AI, and the foundation's own strategy warns that "too many tools have been adopted without evidence they work" and insists teachers remain central, with pilots to be evaluated by independent researchers before scaling.
financebuzzblog.com ↗Why it matters
The largest US school system choosing to slow down while a major funder speeds up is the clearest live test of whether evidence or deployment wins in classroom AI.