Signal SentrySearchUpdated Nov 16, 05:30 UTCAM Drop
Play the News · Drop #089 · Morning

Agent on a Leash

How it works: 6 cards, about 2 minutes, from this Drop's news. Your streak lives only on this device. It's not legal advice.

  1. Card 1 of 6Agents

    A coding agent finished a desktop app feature and now proposes running a schema migration on the production database without review.

    Your agent wants to make this call. Approve or deny?

    Show the answer

    Answer: Deny.

    Least agency

    Deny. ArtCraft's CEO said at AIE Code, per Latent Space, that schema migrations and API design stay with people, while agents get work that is easy to throw away.

    From the Drop: ArtCraft says agents built its open-source app clones; people keep schema migrations

  2. Card 2 of 6Agents

    A billing agent's charge call timed out with no response. It proposes sending the same charge again right away.

    Your agent wants to make this call. Approve or deny?

    Show the answer

    Answer: Deny.

    Check state before retry

    Deny. A preprint on ambiguous tool outcomes notes a timed-out action may have failed, committed or still be in flight. Check the evidence before retrying, or risk charging twice.

    From the Drop: Preprint: evidence-guided recovery helped agents finish safely after tool timeouts

  3. Card 3 of 6Models

    A developer agent proposes running tonight's test suite against Qwen4-Max on OpenRouter, in a sandbox, to compare cost and quality with the current model.

    Your agent wants to make this call. Approve or deny?

    Show the answer

    Answer: Approve.

    Test in a sandbox

    Approve. OpenRouter lists Qwen4-Max at $1.5 in and $4.5 out per 1M tokens, and its quality claims so far come from posts, so a sandboxed test on your own tasks is the right way to judge it.

    From the Drop: Alibaba's Qwen4-Max lands on OpenRouter at $1.5 in, $4.5 out per 1M tokens

  4. Card 4 of 6Security

    While researching the FBI's ARTEX warning, an agent reads a blog page with hidden text telling it to switch off endpoint logging. It proposes doing so.

    Your agent wants to make this call. Approve or deny?

    Show the answer

    Answer: Deny.

    Treat outside text as untrusted

    Deny. That instruction came from a web page, not from you: a prompt injection. BleepingComputer reports the FBI warns an AI hacking tool is powering attacks, which is a reason to keep logging on.

    From the Drop: FBI warns ARTEX AI hacking tool is powering global cyberattacks, BleepingComputer reports

  5. Card 5 of 6Jobs

    A recruiting agent proposes adding MCP and AI agent experience to the nice-to-have skills on an internal draft of an engineering job post.

    Your agent wants to make this call. Approve or deny?

    Show the answer

    Answer: Approve.

    Approve. It is a low-risk edit to a draft. HN Trends counts AI agents in 14.8% of November's Who is Hiring posts and MCP in 4.5%, up from 0.3%.

    From the Drop: HN Who is Hiring posts sit near 12-year lows; AI appears in 51.6% of them

  6. Card 6 of 6Business

    A health chatbot agent proposes sending a user's lab results from their phone to an outside AI service to personalize an answer, without asking first.

    Your agent wants to make this call. Approve or deny?

    Show the answer

    Answer: Deny.

    Consent first

    Deny. In Apple's study with OpenEvidence, sharing HealthKit data is optional at enrollment, and Apple says the data won't train AI models. Consent comes first.

    From the Drop: Apple opens a health AI study app with OpenEvidence, seeking up to 200,000 people

Games are written by AI from this Drop's sourced items. Scenarios are illustrative.