Your AI agent should not hold the keys
Two of today's stories ask the same question: who controls the expensive and risky parts of your product? NVIDIA's OpenShell keeps access rules outside the agent, so a model can't quietly widen its own reach. Anthropic's proof work shows that verifying code can cost hundreds of millions of tokens, so scope your guarantees before you promise them.
- OpenShell gives agents a locked room
- Proving code bug-free gets practical
- Mistral's voice models start at $0.0035/min
- Builders question why OpenAI sells models
- Models repeat the same take on humans
Top stories
All news →What happened: NVIDIA's open-source OpenShell runs each AI agent in its own sandbox, with strict limits on files, network traffic and credentials.
- Every network connection passes a policy check before it leaves the sandbox.
- Agents never see real credentials; OpenShell adds them only to requests bound for approved endpoints.
- Released under Apache 2.0 at GTC on March 16, 2026, and part of NVIDIA Agent Toolkit; reviewers call it alpha software.
Why it matters: Access rules sit outside the agent, so a model that goes wrong can't expand its own permissions. If you sell agents to businesses, expect buyers to ask how you enforce limits like these, and build on tools you can audit.
What happened: Anthropic and Theorem formally verified 239 entry points across 123 open-source libraries in a few weeks.
- The proofs use Rocq, checked by the Rocq kernel, at RISC-V assembly level.
- The output is about 180 million lines of Rocq, and a typical proof runs about 600,000 lines.
- Hundreds of bugs surfaced, plus an unexploitable corner case in about six zlib-style libraries; proofs can take hundreds of millions of tokens.
Why it matters: Token spend per proof decides how much of a codebase you can verify, so start with your security-critical entry points. If you sell code quality, a verified guarantee now looks plausible, but the token bill is the constraint.
What happened: Mistral launched Voxtral 3, a family of speech-to-text and text-to-speech models plus a voice agent SDK built on LiveKit.
- Transcription covers 50+ languages, with diarization (telling speakers apart) and context biasing; realtime runs under 500ms.
- Text-to-speech adapts emotion and pacing, with voice cloning from 30+ minutes of audio.
- Batch transcription starts at $0.0035 per minute through Mistral's API; the launch drew 211 upvotes on Product Hunt.
Why it matters: Low-cost, multilingual voice sets a price floor, so check whether your voice margin still holds. Test it against your current vendor before you commit to a switch.
Quick hits
- heal.dev's Heal 2.0 hunts bugs code review misses, like a discount code that stays applied while checkout charges full price. Read →
- Hacker News asked why OpenAI sells models instead of starting companies; read it as a question about who will compete with you. Read →
- Twelve text models gave nearly the same opinion on humans, so test opinion-style prompts with repeated runs and varied phrasing. Read →
- A 360M model that pretends to be a terminal shows small models can do narrow jobs. Read →
- A researcher argues AI could turn shared science into private property, so watch who controls access to research data. Read →
Deals
All deals →Newsletters & podcasts today
All newsletters →
