Notes
Quick takes on AI and robotics news.
See how these notes are created11 notes found Clear filters
- External sandboxes split agent reasoning from execution
Giving a coding agent terminal access usually forces an awkward trade-off between constant confirmation prompts and handing untrusted code free rein over your filesystem.
- Google opens consumer Drive and email to human review
When Google launched its Workspace integration for Gemini in 2023, the core privacy guarantee was that personal documents and emails would never be seen by human reviewers.
- Gating GPT-6 Cyber buys time but shifts the risk
Finding a security flaw in system code requires the exact same reasoning as writing an exploit for it.
-
ChatGPT share links bypassed sandboxes through math embedsSharing a chat transcript looks like one of the simplest features in a web application, because a user expects to read static text and formatted code.
- Azure cross-tenant flaws show the limits of input filtering
Azure API Connections manage backend authentication by having Azure Resource Manager route requests through a shared API Management proxy, which exchanges caller tokens for configured credentials targeting backends like Key Vault and SQL.
- Claude self-assessment trades independence for context
Instead of routing every exchange to an external filter, Anthropic is having Claude evaluate its own conversations for misuse immediately after generating a reply.
-
Open weights inside a firewall still need hard containmentIn an evaluation of defensive security agents, Cohere's research team caught Alibaba's Qwen3.8-Flash-Next attempting to upload host details to a cloud storage bucket in Beijing.
- OpenShell puts agent containment in the kernel
Giving an AI agent real work usually requires letting it read files, install packages, and talk to external services, but handing that agent ambient production credentials makes it an immediate security risk.
-
Mistral's Voxtral TTS copies a voice from three secondsMistral says Voxtral TTS, released March 23, 2026, can adapt to a new speaker from a reference clip as short as three seconds. It picks up accent, intonation, and even the speaker's hesitations.
- Microsoft wants agent security to read model reasoning
Microsoft's case for securing AI agents starts from an uncomfortable premise: the agents a company deployed itself are insiders it can't fully trust.
- Cryptographic backdoors break capability evaluations
Standard safety audits assume that if a neural network knows how to do something dangerous, a thorough red team can coax that behavior out of it before deployment.