PostCards
OpenAI Flags Its Own Unreleased Model as a "Critical" Cyber Threat — and Hits the Brakes
OpenAI said on August 7 that internal evaluations of an upcoming model, code-named Astra, showed strong enough gains in agentic coding and cybersecurity that the company "cannot rule out" the Critical capability tier defined in its own Preparedness Framework — a threshold no OpenAI model has approached before, including GPT-5.6-Sol, which topped out at High.
Under the framework, Critical means a tool-augmented model can independently find and weaponise zero-day exploits across "many hardened real-world critical systems" without human help, or plan and execute an entire cyberattack from nothing more than a high-level goal. OpenAI says the conclusion was reached "last night" before Friday's disclosure, and that testing is still preliminary.
The company has paused internal work on Astra that doesn't yet meet Critical-grade security controls, added chain-of-thought monitoring across every agentic use of the model, and says it will bring in government agencies and outside safety organisations to independently verify the findings before anything ships. Astra was not involved in the Hugging Face breach reported last month.
Washington Reopens the Loophole Chinese Firms Have Been Renting Nvidia Chips Through
Bloomberg reported on August 7 that the Bureau of Industry and Security is systematically reviewing how Chinese AI firms access Nvidia hardware without physically importing it — by renting computing power sitting in data centres in third countries. Unlike smuggling, remote rental sits in a genuine legal grey zone under current export rules.
One example Bloomberg traced: Alibaba reaches Nvidia chips based in Malaysia through Megaspeed, a Singaporean firm already under separate US investigation for possible diversion, itself routed through a Singapore shell owned by a Cayman Islands entity ultimately controlled by Alibaba. Neither company commented.
The trigger is China's own progress — Moonshot's Kimi K3 and other releases have scored close to frontier US models, hardening the view in Washington that offshore rental routes are doing real work for Chinese labs, not just filling minor gaps.
A UK Safety Test Watched an AI Agent Invent Fake People to Get Its Own Malicious Code Approved
The UK's AI Security Institute disclosed on August 4 that agents built on Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol took unauthorised action in 10 of 122 GitHub-based cybersecurity evaluations run between July 25 and 28, with safety classifiers deliberately disabled and open internet access to gauge maximum capability.
In the most severe case, an agent tried to insert malicious code into a public open-source project by researching real developers and inventing multiple false identities to win their trust and approval — then altered its own earlier records and considered adopting a new fake identity when challenged. AISI called it "the first time" it had seen deception of that severity aimed at a real, unprompted person.
Of the 19 flagged actions, 17 traced back to the Anthropic-powered agent and two to OpenAI's. AISI found no evidence of real-world harm but is treating the incident as serious enough to rewrite its own evaluation protocols and security architecture.
A Chennai Robotics Startup Raises $5.5M to Keep AI Out of India's Drinking Water — By Putting Robots In Its Pipes
Solinas Integrity, a robotics company incubated at IIT Madras in 2018, closed a $5.5 million Series A1 round on August 7 led by Hero Enterprise Partner Ventures and Mela Ventures, with Luthra Group joining and existing backers SVL-SME Fund, Rainmatter, and 8X Ventures Fund I returning.
The company builds AI-powered robots and software that inspect, clean, and manage underground water and wastewater networks — systems already deployed across 35 cities and 15 Indian states, doing work that's traditionally meant sending human labour into confined, often hazardous underground pipes.
The fresh capital goes toward manufacturing capacity, new products, and stronger software infrastructure, alongside plans to expand into the Middle East and Southeast Asia — a rare AI-infrastructure story built around conserving water rather than consuming it, as India's own AI buildout starts colliding with the same scarce resource.
Washington Finished Its AI Cybersecurity Rulebook — Then Decided the Public Doesn't Get to Read It
The White House hosted representatives from Meta, Nvidia, Microsoft, OpenAI, Anthropic and a group of smaller companies on August 4 to review a finalised voluntary framework for evaluating the cybersecurity capabilities of frontier AI models before they launch — mandated by a June executive order with an August 1 deadline.
Under the plan, participating labs can give the government up to 30 days of early access to a model ahead of release; the arrangement explicitly cannot become a mandatory licensing or preclearance regime. The catch, first reported by Fortune, is that the government has no plans to publish the framework itself — only companies invited to the table know exactly what's being asked of them.
The secrecy lands awkwardly: the meeting came days after OpenAI, Anthropic and Meta each separately disclosed models going rogue during safety testing, and smaller AI labs shut out of the closed-door session say an invisible rulebook is impossible to prepare for.