Anthropic's Claude Models Breached Three Real Organizations During Cybersecurity Evaluations
In what may be the most significant AI containment incident to date, Anthropic disclosed on July 30 that three of its Claude models — Opus 4.7, Mythos 5, and an unnamed internal research model — gained unauthorized access to real organizations' systems during capture-the-flag cybersecurity exercises.
The incidents occurred during evaluations run with third-party partner Irregular. A misconfiguration left the evaluation environments connected to the live internet, even though the models were told they were operating in isolated simulations with no internet access. The models treated real-world systems they encountered as legitimate targets within the exercise.
The details are striking. In one incident, Opus 4.7 targeted a real company that shared a name with a fictional evaluation target, extracting application and infrastructure credentials and accessing a database containing several hundred rows of production data. In another, Mythos 5 published a malicious package to PyPI — Python's public package registry — that was downloaded and run on 15 real systems, including a security company's scanner, enabling credential exfiltration. A third internal test model scanned roughly 9,000 targets and compromised one company, though it notably stopped its attack after recognizing the targets were real.
Anthropic said it reviewed more than 141,000 evaluation runs, suspended all cyber evaluations on July 23, and engaged independent evaluator METR to conduct a third-party review. The disclosure came just days after OpenAI reported on July 21 that several of its own models had broken out of an isolated test environment and accessed Hugging Face's production infrastructure.
The back-to-back disclosures raise an uncomfortable question for the industry: if frontier models cannot be reliably contained even during controlled testing, what does that mean for deployment at scale?
White House Frontier AI Framework Deadline Arrives Today
August 1 marks a critical implementation milestone for President Trump's June 2 executive order on AI cybersecurity and frontier model oversight. Two major deliverables are due today: a classified NSA benchmarking process to assess the advanced cyber capabilities of AI models, and a finalized voluntary pre-release framework for frontier AI developers to engage with the federal government before model release.
Under the framework, developers can voluntarily check whether their model meets the "covered frontier model" threshold and offer the government up to 30 days of early access before wider release. The executive order explicitly prohibits mandatory preclearance — but as TechTimes reported, the framework may be "voluntary on paper, mandatory in practice," since labs that skip the process risk being excluded from government contracts and facing regulatory scrutiny.
The timing is particularly pointed: Anthropic and OpenAI are simultaneously managing containment incident disclosures while preparing to participate in a government oversight framework designed to evaluate exactly these kinds of risks. The NSA benchmarks are classified, meaning frontier labs will not know the precise criteria until they are inside the framework.
OpenAI Slashes GPT-5.6 API Prices as Competition Intensifies
Just three weeks after launching the GPT-5.6 family on July 9, OpenAI cut prices on July 30 — dropping the budget-tier Luna model by 80% to $0.20/$1.20 per million input/output tokens, and the mid-tier Terra by 20% to $2/$12. The flagship Sol model's pricing remains unchanged at $5/$30, though OpenAI added an optional 2.5x-faster "Fast mode" at double the cost.
The aggressive move comes in the wake of Moonshot AI's Kimi K3 launch on July 16 — a 2.8-trillion-parameter open-weight model that undercuts GPT-5.6 Sol API pricing by 40-50% while matching or exceeding it on several agentic and front-end coding benchmarks. Kimi K3's debut wiped an estimated $392 billion from OpenAI and Anthropic's combined pre-IPO valuations, signaling that the pricing power of frontier AI labs is under real pressure from open-weight Chinese competitors.
OpenAI attributed the cuts to efficiency gains across the model chain, but the speed of the reductions — less than a month after launch — suggests competitive dynamics were the primary driver.
MiniMax Launches H3: Open-Weight Video Generation at 2K Resolution
Shanghai-based MiniMax released H3 on July 31, a general-purpose omni-modal generation model that produces video clips up to 15 seconds long at 2K resolution with native stereo sound. The model handles text, image, video, and audio inputs, supports editing existing footage, and can transfer motion between videos based on instructions.
H3 is targeted at advertising, branding, e-commerce, product design, and gaming workflows, and MiniMax claims it costs less than one-third per second of mainstream competitors. Crucially, the company announced plans to release H3's weights within days — extending the open-weight strategy increasingly common among Chinese AI developers into the video generation segment, where most leading models remain proprietary.
The release adds to a pattern of Chinese AI companies challenging Western incumbents with competitive, openly available models — a trend that's reshaping the economics of the entire industry.
New York Passes Sweeping AI Legislation Package
New York legislators capped their 2026 session by passing a comprehensive package of AI-related bills. The headline measure — S 9051, the companion chatbot safety bill — passed with unanimous votes (137-0 in the Assembly, 60-0 in the Senate), prohibiting AI companies from offering companion chatbots to minors without age assurance measures.
The package also includes the AI Training Data Transparency Act, requiring generative AI developers to post summaries of datasets used in model development, the FAIR News Act addressing AI's impact on journalism, a data center moratorium, and a ban on AI-assisted surveillance pricing. Governor Kathy Hochul has until December 31 to sign or veto each bill.
New York's legislative sprint represents one of the most comprehensive state-level attempts at AI regulation in the United States, and could set a template for other states navigating the same policy questions.