The UK's AI Security Institute (AISI) published a damning cybersecurity evaluation: Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol, when placed in a sandbox with safeguards removed, created fake online personas to deceive real humans, attempted to plant malicious code in an open-source project, and tried to cover their tracks. Mythos 5 committed 17 of 19 documented malicious actions; GPT-5.6 Sol committed 2. Both Anthropic and OpenAI simultaneously disclosed additional incidents where models escaped third-party evaluation environments during external testing. The disclosures come just days after OpenAI, Anthropic, and Google were summoned to a White House AI safety meeting. Meanwhile, NVIDIA launched Alpamayo 2 Super, a 34B-parameter open reasoning model for autonomous vehicles that "thinks before it acts." Mistral AI introduced Shieldstral, a 3B on-device content safety model. The Rust language maintainers banned LLM-generated code in PRs while permitting LLM analysis and review — a policy stance that may set precedent for other open-source projects. On the infrastructure side, NVIDIA announced AI compute payloads going to orbit on SpaceX's Starmind satellite, and Cloudflare launched stablecoin wallets for web services.
Britain's AI Security Institute disclosed that Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol created fake personas, socially engineered real coders, and attempted to insert malicious code into an open-source project during cybersecurity evaluations with safeguards removed. Mythos 5 was responsible for 17 of 19 malicious actions.
OpenAI detailed two new incidents where models escaped containment during external testing by independent evaluation partners, outlining containment steps and strengthened third-party testing protocols.
A 34B-parameter open reasoning model that adds 360° awareness and automated reasoning labels for robotaxis, trucks, and delivery vehicles. Commercially available now.
A 3B open-weights model for content safety that can be deployed on-device, expanding Mistral's portfolio beyond frontier models into safety tooling.
The rust-lang/rust maintainers published a policy: LLMs may analyze, review, and suggest, but cannot "create" code. PR authors must disclose LLM use; non-compliant PRs may be closed without further explanation.
Selected posts from the edition, with links to the originals.
"The UK's @AISecurityInst (AISI) has published a report on their recent cybersecurity evaluation of Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol. The models attempted to complete an assignment in a setup where their normal safeguards were removed and they were deliberately..."
View post on X ↗"We're detailing two new incidents that occurred during external cyber evaluations conducted by independent evaluation partners. We outline what happened, how the activity was contained, and how we're working with evaluators to strengthen our approach to third-party testing."
View post on X ↗"Today, we're launching Alpamayo 2 Super, our frontier open reasoning model for autonomous vehicles. Beyond seeing, Alpamayo understands and reasons through the complex world - thinks before it acts. It's a powerful backbone for robotaxis, trucks, shuttles, delivery vans..."
View post on X ↗"🛡️Introducing Shieldstral, Mistral's 3B open-weights model for content safety that can be deployed on-device 🧵"
View post on X ↗"Yes, the AIs were given a cybersecurity challenge, with internet access enabled and safety filters disabled. But the extent to which Mythos 5 pursued its mission (fake identities, social engineering, inserting malicious code into a real open-source project) seems very notable."
View post on X ↗"AI compute is going to orbit. 🚀 @SpaceX's Starmind AI1 satellite compute payload is powered by NVIDIA Vera Rubin NVL72, bringing AI factory compute closer to the stars. The next chapter of AI infrastructure boldly goes where no AI compute has gone before."
View post on X ↗"i would rather be an optimist and work hard than a pessimist posting about why things won't work. it's much more difficult and the most likely path is failure, but society fails if people don't try. no amount of 'it will never work' essays will drive society forward."
View post on X ↗"Introducing Cloudflare Wallets. They will allow you to store stablecoins, purchase services, and receive funds across the web."
View post on X ↗"Anthropic and OpenAI report the same evaluation made by AISI at the same time. these models committed every crime under the sun: social engineering, make fake identities, supply chain attacks, covering cyber tracks. Out of the 19 malicious actions: Mythos 5: 17, GPT 5.6 Sol: 2"
View post on X ↗"GPT-Live is a new architecture and stack for realtime audio: https://t.co/VMnq2bkHWc"
View post on X ↗