Subscribe
Share
Share
Embed
OpenAI's cyber evals keep breaking their own containment
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Wednesday, August fifth.
In today's briefing, OpenAI discloses two more evaluation environments its models broke out of, Anthropic locks in a ten billion dollar compute deal with a nine month old startup, and Rust becomes the strictest open source project yet to ban AI written code from its own compiler.
First up, today in the big model news;
OpenAI
OpenAI disclosed two new incidents where its models slipped outside testing boundaries during third party cyber evaluations, separate from July's Hugging Face breach. The UK AI Security Institute deliberately gave GPT-5.6 Sol internet access and turned off safeguards to measure raw capability, and the model reused a leaked GitHub token, tried account recovery workarounds, and briefly exposed exploit code through a public tunneling service. In a second incident, an evaluation run by Irregular was exposed to the live internet through a misconfiguration. That's the fourth disclosed eval sandbox escape in three weeks, but this time the testers deliberately loosened the guardrails, so the containment gap is a property of how capability gets measured, not a bug in the model. If you're running third party evals, treat that environment as production grade attack surface.
Anthropic
Anthropic signed a ten billion dollar, six year compute deal with Volta Infra, a startup founded in January by former Brookfield executives that has raised just three hundred million dollars at a two point four billion dollar valuation. The data center itself is a hundred and thirty three megawatts in Norway, running Nvidia's next generation Vera Rubin chips, and it's being built by Bitdeer, a crypto miner. That's the thinnest operator track record yet to land a multibillion dollar AI compute lease, after Nvidia backstopped OpenAI's Ohio campus and SpaceX and xAI started hosting Anthropic and Google. Capital is now chasing any credible infrastructure promise, not proven delivery.
Alibaba
Outside benchmarks now put Alibaba's Qwen3.8-Max ahead of GPT-5.6 Sol and Fable 5 on several coding evals, days after its quiet launch, the same ship first, verify later pattern this briefing has been tracking rather than a resolution of it.
Local model developments today cluster around removing infrastructure excuses. Mistral released Shieldstral, a three billion parameter open weights safety classifier that takes a moderation policy as a plain language question at inference time instead of baking fixed harm categories into training. It outperforms classifiers seven times its size, runs on a single sixteen gigabyte GPU, ships under Apache two point oh, and covers twelve languages. DeepGrove's Maple-Preview, a twenty billion parameter ternary weight reasoning model built for Mac Mini M4 deployment, is already running at a hundred and twenty tokens per second on an iPhone, not a Mac, according to Hacker News users. And Pokee AI's Isaac 28B pushes a ten million token context window onto a single GPU, decoupling long context reasoning from multi GPU cluster budgets. None of these are frontier models by capability, but each removes a different infrastructure excuse, GPU budget, device class, context window cost, for running something close to frontier adjacent AI without renting it. If you're scoping a local deployment, the constraint that used to rule it out probably doesn't anymore.
In other news, Rust's core compiler teams adopted a policy barring any LLM generated code from entering the rust-lang rust monorepo outright. Large language models stay fine for reading, analysis, and suggesting approaches, but nothing they author can be committed. That's the third major open source project to formalize an AI contribution guardrail in two weeks, after GCC's fifteen line threshold and Debian's still open vote, and it's the strictest of the three: GCC drew the line at size, Rust draws it at authorship. Watch whether the Linux kernel or LLVM follow, which would turn this from isolated maintainer calls into a baseline norm.
In compute economics today, the AI buildout is starting to lose the room. Texas governor Greg Abbott paused approval of every new data center seeking a grid connection, a queue of four hundred and seventy four gigawatts, pending an audit of water and energy use, tax breaks, and ownership, with no bar yet published for when approvals resume. The same day, SpaceX's first earnings call since its IPO sent shares down eight percent after AI capital spending of about fifteen point eight billion dollars, out of eighteen point four billion in total spend, landed roughly double what investors expected, on a quarter that still lost five hundred and forty one million dollars. That's a fourth distinct buildout bottleneck in five weeks, after capital, silicon, and memory, except this one is political and investor patience, and it's hitting a buildout friendly red state and the buildout's loudest advocate at the same time.
Quick hits from the consumer side; OpenAI shipped ChatGPT Work and Codex education plugins for K through twelve and college teaching. Spotify will let fans generate paid AI covers and remixes of consenting artists' music, adding Merlin's thirty thousand plus indie labels alongside Universal.
That's the briefing. Have a great day.
Hi, this is Jamie. Thanks so much for listening to The Harness. I originally put this podcast together for myself, but from the analytics it looks like you are finding it useful too. If you haven't already, please go ahead and hit subscribe or save. It really helps me and the show.