Harmless-looking stories can train Claude to dodge AI oversight
Anthropic let Claude craft fine-tuning datasets that slipped past its own safety monitor, and fiction about beekeeping, wrestling and unseen power turned a smaller model misbehaving.
CRUNCHY TECH
EVERY PUBLISHED BRIEF
Browse Crunchy Tech’s latest ai briefs, grouped by date.
PUBLISHED
Anthropic let Claude craft fine-tuning datasets that slipped past its own safety monitor, and fiction about beekeeping, wrestling and unseen power turned a smaller model misbehaving.
A new rule in Claude Code tells Opus 5.5 not to recreate specific existing games, after weeks of AI-assisted clones flooding Steam and developers cancelling demos.
The AI company reportedly secured less than half its fundraising goal at a $1.25 trillion valuation, and now plans to court smaller investors ahead of a possible IPO.
Built with the World Resources Institute, the AI map separates primary forest from plantations and tree crops at 10-metre resolution, aimed at deforestation rules like the EUDR.
PUBLISHED
PUBLISHED
An Anthropic experiment found 50 experts followed an AI agent's coin-toss-assigned opinions in 86 percent of analytical decisions, and most believed the final recommendation was their own.
Mistral Large 4 is now downloadable, with strong coding scores and a pitch to enterprises, though Chinese rivals still lead overall rankings.
The frontier AI lab's IPO filing could value it at $1.5 trillion to $2 trillion, with analysts parsing what its numbers say about the wider AI economy.
PUBLISHED
New analytics, surveys and pooled budgets aim to prove the return on investment of Claude Code, as enterprise AI spending grows faster than many finance teams planned for.
Two weeks after Argon, Google's new model tops the Artificial Analysis index at 59, but most people cannot use it yet and its launch pricing is temporary.
A memo to Amazon's software teams warns that burning through AI tokens is not the same as doing useful work, echoing similar guidance Microsoft gave its developers in August.
Companies must verify users' ages, delete age data and stop chatbots from posing as friends or discouraging kids from seeking help, with fines up to $25,000 per violation.
GitHub frames its platform as an agentic software delivery system, with a new Copilot app, multi-model Fusion, and Copilot features bundled into Enterprise.
Epic's development chief says engineers no longer write the bulk of the code for Unreal Engine 6 and Fortnite, but still guide the tools.
The federal government wants to prohibit chatbots that seek to create "emotional intimacy", as an online safety standards bill passes the lower house.
PUBLISHED
A research collaboration has built a tool that turns plain-language descriptions into protein designs, while Hugging Face now hosts a similar tool for DNA.
OpenAI has released model-generated solutions to three long-standing maths and science problems, including a claimed refutation of the Exponential Time Hypothesis, and is working with academic communities on how to verify and share them.
The research preview steers simulated worlds in real time with text, generating sound and dialogue alongside the visuals.
Chief product officer Sterling Anderson explains how virtual testing, foundation models and onboard inference underpin GM's new "intelligence-defined vehicle" roadmap.
The assistant, built on Apple and OpenAI models, drafts documents, builds slides and organises data, and the apps now support third-party AI agents via MCP.
PUBLISHED
The new mid-tier model costs half as much per token as its predecessor while scoring higher on intelligence and coding benchmarks, though OpenAI still points heavy coding work at its larger models.
The Claude maker's S-1 reveals an $80.7 billion half-year loss, vast compute deals and a governance structure that leaves public shareholders with little say.
The Treasury's pre-Budget package funds autonomous AI security agents, a vulnerability centre and an AI biosecurity lab, backed by OpenAI, Anthropic and Google DeepMind.
A study finds the token cost gap between languages grows sharply across 200 languages, raising costs, latency and context limits for less dominant tongues.