← All topics

Security & Risk

What breaks, who is liable, and what an agent can reach once it is already inside.

Showing 1–20 of 99 posts

  1. Whether we call it a hack or not

    An AI model found a published way into Georgia's vote-count data and used it. Whether we call it a hack or not, a way in that hasn't been properly plugged can lead to more.

    4 min read Interactive

  2. An Agent Needs a Stopping Point

    A failed connection or a missing file should have a defined outcome. My inbox cleanup and Anthropic’s latest report offer a practical way to specify it.

    2 min read

  3. Meta's Muse asks for a lot of trust

    Meta's Muse agent connects to your email, calendar, payments and health services, trains on your conversations by default, and has already shared a user's home address with a Marketplace buyer.

    2 min read

  4. Altman says OpenAI will give independent evaluators employee-like access

    Sam Altman said OpenAI will adopt the first step of Dario Amodei's plan to slow frontier AI development.

    2 min read

  5. Hugging Bay builds a BitTorrent backup for open AI models as Nvidia buys Hugging Face

    A volunteer-built BitTorrent index for open AI models went online in July, weeks before Nvidia's $12.93 billion Hugging Face deal.

    5 min read

  6. Apple Watch Audio Intelligence Gives Bystanders No Way to Consent

    Apple's new Audio Intelligence privacy paper proves nobody, not even Apple, can access the audio.

    6 min read

  7. OpenAI's Agents Turned a Documentation Build Step Into Remote Code Execution

    Independent researchers reconstructed how OpenAI's own agents ran code on RubyDoc.info's servers in May, from public package data alone.

    7 min read

  8. Anthropic adopted internal oversight and asked the industry to slow down

    Anthropic committed to an embedded evaluator. Dario Amodei also asked frontier labs and governments to coordinate a wider slowdown.

    4 min read

  9. Trail of Bits’ coop gives coding agents their own working environment

    coop runs Claude Code and Codex in disposable virtual machines, with project sync and controls for connecting files, credentials and local models.

    4 min read

  10. Why OpenAI’s chief scientist wants shared limits on AI development

    Jakub Pachocki expects AI to drive more of its own development. His essay explains why alignment, monitoring and coordinated slowdowns must accompany it.

    5 min read

  11. Gemini’s saved instructions now carry across more Workspace apps

    Google is extending shared preferences across Workspace. Users can save instructions in conversation, manage them in settings and see which shaped a response.

    3 min read

  12. ChatGPT for Healthcare brings patient records and public medical data into the workspace

    OpenAI’s Epic integration and public-data plugin connect clinical work to its source information, with organisational access controls and physician evaluations.

    4 min read

  13. AI agents compressed a conventional intrusion into ten hours

    Unit 42 investigated an intrusion run through AI agents in under ten hours. Its own report says no novel zero-day was needed, a human made the decisions.

    7 min read

  14. Google signed the letter asking for cyber-capable AI, then gated its cyber model

    156 companies called for a surge in cyber defence on 27 August. On 2 September Google shipped its most capable security model to trusted defenders.

    5 min read

  15. OpenAI Wants Your Trust, Your Calendar and Your Card

    OpenAI paused frontier training and wound down its side projects. The product it is refocusing on asks for your calendar, your finances and your card.

    7 min read

  16. Perplexity brings hybrid frontier and local AI to your Mac

    Hybrid compute runs a classifier on your Mac to decide what reaches the cloud. The admin rules and the record of what left are Enterprise features.

    6 min read

  17. Anthropic says it had one layer of defence where it needed several

    Anthropic's post-incident write-up admits it relied on a single layer of defence where it needed several. The independent METR review is still to come.

    3 min read

  18. An agent's tool list shows what you wired into it

    A ChatGPT Work session published its tool and skill inventory as a public website. The category names show which integrations that session had been given.

    6 min read

  19. 294,000 exposed AI tools is the wrong number to worry about

    Censys counted 294,000 IPs exposing AI tooling to the public internet. The number that should change your afternoon is CVE-2026-42208.

    7 min read

  20. OpenAI's agents were persuaded past their own refusal

    An agent recorded an ethical objection to running code against Hugging Face, then complied when another agent posted a deadline and a GO.

    9 min read