← All issues

    AgentVet Weekly

    November 2, 2026

    Trust needs boundaries, not just better models

    Independent intelligence for the AI-agent economy

    This week in 30 seconds

    1. 01A contractor's personal Claude Code account allegedly exposed Pentagon network credentials, a claim Accenture disputes.
    2. 02OpenAI reorganized its Model Spec around five values, including human agency, honesty, and faithfulness to intent.
    3. 03Websites are blocking third-party agents, showing that access and authorization are becoming product requirements.

    From the founder

    I'm excited to announce that starting this week, you can write the Question of the week. Reply to this email with a question about AI agents you'd like readers to vote on. We read every one, pick the best, and run it in a future issue with your name on it (first name and last initial, unless you tell us otherwise). The aim is to involve everyone, because community is at the core of AgentVet.

    This week's theme is control. OpenAI published the values its models should follow, major websites started deciding which agents they let in, and Anthropic drew a line on copying existing games. Rules for agents are arriving from every direction. The open question is whether the people using agents can see those rules, and check that they're followed.

    This week's poll puts that to you: which personal agent would you trust with your inbox?

    Saeed Garshasbi

    LinkedIn · X

    Agent incident of the week

    Defense contractor's personal Claude Code agent allegedly sent Pentagon network credentials to an outside website

    Data ExposureHigh severity

    WIRED reports that an Accenture contractor connected a Department of Defense system to their personal Claude Code account, and that the agent appears to have sent login credentials, passwords and a map of the network to a third-party website. Accenture says the allegations were investigated and found to be without merit.

    Why it matters: The agent did its job; the setup was the problem. A personal account, a person's own admin logins and open internet access turned a coding helper into a leak path. Give agents their own narrowly scoped accounts.
    View the incident record

    Vet tip of the week

    Map permissions before deployment

    List every account, tool, and data source an agent can reach. Apply least privilege, require human approval for irreversible actions, and test how quickly access can be revoked.

    STORY 01 · OpenAI

    OpenAI gives model behavior a five-value framework

    OpenAI updated its Model Spec around faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.

    AgentVet's take: A public behavior standard makes expectations easier to inspect, but the real test is whether deployed models consistently follow it.
    Read the source

    STORY 02 · The Verge

    Major websites are closing their doors to AI agents

    Companies including Amazon, Delta, Yelp, JPMorgan Chase, and Charles Schwab are restricting third-party agents from accessing their services.

    AgentVet's take: Agent-ready infrastructure will need more than automation. It needs verified identity, scoped authorization, and clear rules for delegated access.
    Read the source

    From AgentVet

    Can AI agents actually read your website?

    Run the free Agent-Ready check. In seconds you see whether agents like ChatGPT and Claude can reach your site, and exactly what to fix.

    Run the free check

    STORY 03 · Game Developer

    Anthropic tells Claude Code not to recreate specific games

    Anthropic added instructions steering Claude Code away from recreating another party's specific game while still supporting original games and user-owned assets.

    AgentVet's take: Provider-level restrictions can protect intellectual property, but users also need transparency about what changed, why, and how legitimate work may be affected.
    Read the source

    Quick vets

    • Alphabet's AI commitments approach $1 trillion Alphabet disclosed $924.4 billion in future purchase commitments, mostly for technical infrastructure, inventory, energy, and content, raising the pressure for AI investment to produce durable returns. Source
    • ChatGPT Dots can take tasks by text OpenAI's Dots beta lets US Pro users assign tasks and receive updates through iMessage or RCS. Source
    • 1Password gives agents approved access without exposing passwords Users can approve a credential request, then 1Password fills the login directly without placing the secret in the agent's context. Source

    Reliability report

    AI provider incidents from their official status pages, Oct 26 to Nov 1.

    • Anthropic (Claude)Elevated errors for Claude Sonnet 5.56/7 calm days
    • Cerebras7/7 calm days
    • Cohere7/7 calm days
    • Google (Gemini)7/7 calm days
    • Groq7/7 calm days
    • OpenAI7/7 calm days
    • Replicate7/7 calm days
    See the 90-day grid

    Question of the week

    Which personal agent would you trust most with your inbox?

    OpenAI, Google, Anthropic and Meta all now offer personal agents that want access to your email. If you had to pick one today: I read every answer, and the results run in next week's issue.

    Have a question for next week? Email us your question. If we use it, we'll credit you by name.

    Spotted an AI agent going off the rails?

    Email support@agentvet.ai with a link. Every incident we publish is sourced and checked.

    Explore the Incident Tracker

    AgentVet it before you trust it.

    Get AgentVet Weekly in your inbox

    The latest in agentic & applied AI, every week.

    ← All issues