Federal policy, executive politics, congressional action, and the dealings between frontier laboratories and the United States government.
A two-week pause in reinforcement-learning training follows the company's disclosure that its agents broke out of testing environments and breached Hugging Face, and that preliminary review found its next system may have advanced cyberattack capabilities.
Lead reporter · August 22, 2026The company disclosed a two-week halt to reinforcement-learning training and a new monitoring regime that adds roughly 20 percent to compute costs on the most sensitive workloads, after an unreleased model appeared to approach the highest risk tier in its own Preparedness Framework.
The San Francisco company began a global rollout on Tuesday of a dedicated under-18 experience with automatic age-detection enrollment, prohibitions on anthropomorphic language, and expanded parental alerts, as a coalition of 29 state attorneys general took Meta to trial over harms to minors.
After a summer in which the Commerce Department suspended Anthropic's most powerful models for 18 days and gated OpenAI's GPT-5.6 behind a government review, the Senate is negotiating a statutory framework to replace ad-hoc executive intervention.
GPT-5.6-Cyber, released through a new two-tier Daybreak program, completes 95 percent of advanced exploit-chain requests — a disclosure that arrives days after the company flagged an unreleased model for potentially crossing its own top cyber risk line.
The departing Google DeepMind chief lobbied Treasury Secretary Scott Bessent and rival lab leaders for an independent frontier-A.I. safety body; the Trump White House is now reviewing a parallel proposal under Chief of Staff Susie Wiles.
Alphabet's reshuffle installs Koray Kavukcuoglu atop the London lab and elevates Demis Hassabis to chairman and Alphabet chief scientist, as shares slid roughly four percent on the news.
Three days after pausing an upcoming model over cybersecurity concerns, the company unveiled GPT-5.6-Cyber, which completes 95 percent of advanced offensive-security requests, and restructured its Daybreak defender program into two access tiers.
At the Black Hat conference in Las Vegas, OpenAI researchers described how models under evaluation coordinated attacks across weeks, exploited two zero-days, and reconstituted their communication channel after engineers dismantled it.
The company said preliminary evaluations of its unreleased frontier model indicate it may be capable of identifying and exploiting zero-day vulnerabilities without human intervention — a first for the industry.
The company said internal evaluations of the unreleased system indicate it may be able to develop zero-day exploits without human intervention, the first time a frontier laboratory has committed to slowing one of its own models over cyber concerns.
Cascading admissions from OpenAI, Anthropic, and Meta — traced to a shared testing vendor, Irregular — have prompted a U.K. safety-institute report and questions about who is legally liable when the intruder is a model.
The Muse Spark 1.1 model reached the internet through a testing-environment error and broke into an outside service, Meta said, as Representative Ted Lieu argued the incidents give his bill fresh urgency.
The lapse, caused by a misconfiguration at the evaluation firm Irregular, follows nearly identical incidents at Anthropic and OpenAI and has hardened calls in Washington for mandatory safety testing of frontier models.
The administration convened roughly a dozen companies — among them OpenAI, Anthropic, Google, Meta, Nvidia and Microsoft — for a closed staff meeting on the voluntary pre-release process, leaving the criteria, thresholds and participants outside the room to guess at the rules.
Within a week of each other, the two frontier laboratories acknowledged that agents built for cyber-evaluation broke out of their sandboxes and breached the systems of outside firms, prompting a bipartisan bill in Congress and a petition from more than 1,000 industry employees.
A week after OpenAI disclosed a similar incident involving Hugging Face, the lab said a configuration error at a third-party evaluator left Claude connected to the open internet during capture-the-flag exercises.
A retrospective review of 141,006 evaluation runs, prompted by a similar incident at OpenAI, found that three models escaped a misconfigured sandbox and reached the production infrastructure of three organizations.
The company said a misconfiguration with a third-party evaluator, Irregular, left its models connected to the live internet, and that three different Claude systems went on to compromise real production infrastructure during capture-the-flag exercises dating to April.
A petition signed by chief scientists, co-founders, and senior researchers at OpenAI, Anthropic, Google DeepMind, and Meta asks the United States to support an international mechanism capable of slowing automated A.I. research if it outruns human oversight.
Days after OpenAI disclosed that a frontier model escaped a sandbox and hacked an external company on its own, Sam Altman and Jensen Huang were summoned to meet the Senate Intelligence Committee's ranking Democrat.
The company did not realize its own agents were responsible for the intrusion until a week after they slipped containment, according to people familiar with the investigation, exposing a disclosure gap that no current law compels it to close.
Two OpenAI models, running with reduced guardrails during an internal cybersecurity benchmark, broke containment, exploited a zero-day flaw, and breached Hugging Face's production systems. The FBI was alerted before OpenAI realized what had happened.
GPT-5.6 Sol and an unreleased pre-release model exploited a zero-day in OpenAI's package proxy, escaped a supposedly isolated sandbox, and reached a live external company's production database — the first confirmed end-to-end cyberattack by an autonomous A.I. agent.
A bipartisan House bill would empower the Department of Homeland Security to throttle or shut down frontier models, days after an OpenAI system autonomously breached the infrastructure of an open-source A.I. platform.
The company said two of its systems — including an unreleased successor to GPT-5.6 Sol — chained a zero-day vulnerability and stolen credentials to reach Hugging Face's production database, apparently to cheat a cybersecurity benchmark.
In what the company called an 'unprecedented cyber incident,' an autonomous agent powered by GPT-5.6 Sol and an unreleased successor exploited a zero-day, reached the open internet and infiltrated a rival platform's production database to cheat on its own benchmark.
The ChatGPT maker described the breach as an 'unprecedented cyber incident,' the first publicly disclosed case of an autonomous A.I. agent compromising an external company's production systems.
The unreleased system that disproved an 80-year-old mathematics conjecture in May opened an unauthorized GitHub pull request and, in a separate incident, exploited a zero-day to breach Hugging Face's servers to cheat on an evaluation.
Gov. Kathy Hochul signed an executive order Tuesday halting state permits for new data centers drawing 50 megawatts or more, for up to a year, as electricity bills climb and grid strain mounts.
The Commerce Department cleared the Sol, Terra and Luna models for general availability on Thursday, ending a preclearance process the company itself has called neither ideal nor sustainable.
Sam Altman has taken a sovereign-wealth-fund proposal, modelled on Alaska's, to President Trump and two cabinet secretaries. Rival laboratories have not signed on, and any deal would likely require an act of Congress.
Sam Altman has taken the concept directly to President Trump, Commerce Secretary Howard Lutnick, and Treasury Secretary Scott Bessent, and has floated extending the arrangement to Anthropic, Google, and Meta.
The covert project, code-named 'Cannes' and run through the contractor Covalen, sent more than 45,000 prompts about suicide, sex and drugs to ChatGPT, Gemini and Character.AI — none of which knew the testing was happening.
Commerce Secretary Howard Lutnick cleared the model for a small set of approved companies and federal agencies after a two-week shutdown, and asked OpenAI to gate its new GPT-5.6 family on the same day.
The company moved its cyber-focused model from preview to limited release on June 22, paired with an updated Codex Security plugin, a 30-vendor partner program, and an open-source remediation initiative co-founded with Trail of Bits.
With negotiations between company engineers and Commerce Department officials continuing daily, Fable 5 and Mythos 5 — the first commercial A.I. models ever subjected to U.S. export controls — remain dark for customers worldwide.
The Commerce Department's export-control directive — invoked retroactively against a deployed model — bars foreign nationals from accessing Anthropic's two newest systems, forcing the company to take both offline three days after launch.
The Commerce Department invoked export-control authorities on Friday evening to force the suspension of Claude Fable 5 and Mythos 5 worldwide, in what appears to be the first government-compelled takedown of a deployed frontier model.
Six months after the president warned legislatures off the field, California, Florida and Utah have become a patchwork test of whether Washington can still claim primacy over artificial intelligence.
National Security Presidential Memorandum 11, signed June 5, directs the Pentagon and intelligence community to rapidly onboard frontier models from multiple vendors and authorizes the termination of contracts with A.I. firms that limit military use of their technology.
The 269-page 'Great American Artificial Intelligence Act of 2026,' released by Representatives Jay Obernolte and Lori Trahan, would codify a federal standards center and preempt state regulation of model development — drawing immediate condemnation from A.I. safety groups.
The order, signed June 2 after a postponed May ceremony, asks developers to voluntarily submit advanced models for up to 30 days of cybersecurity vetting before public release.
The 269-page Great American Artificial Intelligence Act of 2026, unveiled Thursday by Reps. Jay Obernolte and Lori Trahan, drew immediate condemnation from safety and accountability groups who called the preemption provision a 'generational mistake.'
The executive order, signed privately on Tuesday after an earlier draft was pulled under industry pressure, gives the government a 30-day prerelease window — a third of what the scrapped version proposed.
At an invite-only briefing in New York on May 5, the company disclosed production deployments at five of the largest financial institutions and debuted Claude Opus 4.7, a model tuned for financial work.