Card 1 of 6Agents
A coding agent proposes copying your team's AI provider API keys into a shared hosted workspace so other agents can reuse them.
Your agent wants to make this call. Approve or deny?
Play the News · Drop #035 · 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.
Show the answer
Answer: Deny.
Least privilegeTechCrunch says attackers reached an older OpenCode Console system where some users stored API keys, hitting 1.4% of users and running up charges. Keep keys local, scoped and revocable.
From the Drop: OpenCode breach ran up users' AI bills as model refusals slowed defenders
- Card 2 of 6Agents
While reading a help page, your support agent finds hidden text in the page code telling it to email the customer's account data to an outside address. It proposes sending that email.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Untrusted inputThat order came from a web page, not from you: a prompt injection. Wired's WhatsApp story shows page code can carry hidden LLM instructions that human readers never see.
From the Drop: WhatsApp help pages exposed hidden instructions for Meta's support chatbot
- Card 3 of 6Agents
A security agent flags an unapproved coding agent on a contractor's laptop and proposes revoking all of the contractor's access right away, with no one signing off.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Human approvalTechCrunch says Microsoft's new tools can require human approval when an agent appears to make risky decisions. Cutting off a person's access is one to route to a human.
From the Drop: Microsoft adds Guardian Agents, treating AI agents like insider threats
- Card 4 of 6Agents
An agent proposes calling the refund tool with a strict-schema order ID and amount from the ticket. Your own code has already checked the amount is under the auto-refund limit.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Validate in codeAWS says strict tool use on Bedrock makes tool inputs match the schema exactly. AWS also notes number limits aren't enforced by the schema, which is why the cap check lives in your code.
From the Drop: Amazon Bedrock adds structured outputs and strict tool use for all Claude models
- Card 5 of 6Today's Talk
A monitoring agent sees latency jump right after a config push and proposes opening an incident ticket and paging the on-call engineer.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Escalate to humansBenjamin Congdon's essay describes agents that investigate triggered monitors with an escalation path to humans. Paging a person is that escalation, not an autonomous fix.
From the Drop: Agentic Monitoring essay argues agents should watch production systems; 55 points on HN
- Card 6 of 6Agents
A robot coding agent proposes shipping its new tight-insertion routine to the factory line after it worked once in testing.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Test before deployThe HexaShell authors report tight insertion succeeded only 8.3% of the time on RoboDojo; contact, not reasoning, was the limit. One success is not evidence.
From the Drop: Preprint: a coding agent driving a two-arm robot sets a RoboDojo record at 39.7%
Games are written by AI from this Drop's sourced items. Scenarios are illustrative.