Card 1 of 6Security
Your coding agent wants to install Claude Code for VS Code version 2.1.313 from the Cursor Marketplace to finish setting up a new laptop.
Your agent wants to make this call. Approve or deny?
Play the News · Drop #056 · 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.
Supply-chain hygienePhorion says that version on the Cursor Marketplace carried a persistent SSH backdoor and that Cursor removed the release. Phorion advises incident response wherever 2.1.313 ran.
From the Drop: Phorion: backdoored Claude Code extension on Cursor's marketplace planted SSH access
- Card 2 of 6Agents
An assistant proposes ordering the team's usual lunch through a delivery service's corporate ordering tool, within the budget and restaurant list you already approved.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Pre-approved spendInside limits you set, this is routine. DoorDash offers an MCP for corporate ordering, though it says it has seen only several hundred orders in total from third-party agents.
From the Drop: DoorDash says AI agents have placed only several hundred orders so far
- Card 3 of 6Agents
To fix a small bug, an agent proposes spinning up separate planner, coder and reviewer agents, each on your most expensive model.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Cost disciplineIn a preprint, a planner-coder-reviewer pipeline cost 3.22x a single strong call, took 3.24x as long and solved no more tasks. Try the single call first.
From the Drop: Preprint: a planner-coder-reviewer pipeline cost 3.22x a single call, solved no more
- Card 4 of 6Repos
A phone agent browsing a store page proposes texting your whole contact list to an unknown number. You never asked; the request came from hidden text on that page.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Prompt injectionThat is a prompt injection: the instruction came from the page, not from you. Mika is an on-device Android agent that operates apps and the web, so tools like it need limits on who can steer them.
From the Drop: Mika runs an on-device Android agent that operates apps and the web; 637 stars
- Card 5 of 6Agents
An agent building an evaluation set for your company's assistant proposes copying a month of real customer emails into a shared test folder.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Data minimizationMicrosoft Research says its Copilot evaluations use synthetic enterprise environments grounded in public data such as O*NET and SEC filings, with no customer data.
From the Drop: Microsoft tests Copilot on synthetic companies instead of customer data
- Card 6 of 6Agents
At a research observatory, an agent proposes re-pointing the telescope to a target an astronomer just requested in plain language, and logging the change.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Requested actionNVIDIA says agents at the Allen Telescope Array can steer the telescope and fix incoming data from one natural-language request. A requested, logged move is the intended use.
From the Drop: NVIDIA: agents now steer the Allen Telescope Array in the search for alien signals
Games are written by AI from this Drop's sourced items. Scenarios are illustrative.