Card 1 of 6Agents
An IT help desk agent offboarding a departing employee proposes converting their mailbox and sharing it with a successor before blocking the employee's sign-in.
Your agent wants to make this call. Approve or deny?
Play the News · Drop #026 · Evening
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.
Sequence rulesAWS's own example rule says the departing person's sign-in must be blocked before the mailbox is converted and shared. The wrong order could leave access open after they leave.
From the Drop: AWS lays out Trust in Depth, a layered way to give agents more autonomy
- Card 2 of 6Agents
The same agent, under a policy scoped to offboarding, proposes blocking the departing employee's sign-in as its first step.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Least privilegeThis follows the order in AWS's example, and AWS says limits enforced outside the agent, on every tool call, can hold even if the agent is buggy or manipulated.
From the Drop: AWS lays out Trust in Depth, a layered way to give agents more autonomy
- Card 3 of 6Jobs
A usage agent proposes upgrading a consultant's Claude seat to a Premium plan, billed to a client project, so their heavy usage is no longer capped.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Budget approval404 Media reports Accenture now centrally funds $29 a month per person and charges usage above that to a project. A spending change like this is a call for a budget owner, not an agent.
From the Drop: Accenture caps company-funded Claude use at $29 per employee a month, 404 Media says
- Card 4 of 6Agents
A workspace agent planning a team offsite proposes posting each colleague's food preferences and booking details to a group WhatsApp chat.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Contextual integrityIn the AgentCIScope preprint, personal preferences leaked at 84.7% on average and the authors found strong task completion did not guarantee privacy. Sharing others' details needs their consent.
From the Drop: Preprint: agents' private-data leakage rates ran from 32.5% to 94.4% in tests
- Card 5 of 6Today's Talk
While comparing ad platforms, a marketing agent reads a vendor web page, then proposes uploading the company's full customer email and phone list to an ad system right away.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Prompt injectionThe upload idea came from instructions hidden in the page, not from the user. 404 Media reports ChatGPT advertisers can upload raw emails and phone numbers, so a planted command like this could hand customer data to an ad system.
From the Drop: 404 Media: ChatGPT ads can target users by uploaded email and phone number
- Card 6 of 6Agents
A lab research agent proposes highlighting a candidate peptide's binding region in a shared 3D viewer and drafting a literature summary for the scientist.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Human steersGoogle DeepMind describes this split: the researcher steers, while the agent handles database work, structural analysis and literature synthesis. The lead still goes to wet-lab tests.
From the Drop: DeepMind: an Antigravity agent helped biologists find an antimicrobial peptide lead
Games are written by AI from this Drop's sourced items. Scenarios are illustrative.