Play the News · Drop #050 · 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.

Card 1 of 6Agents
You asked your agent to refresh your own quarterly deck. Through the app's MCP connection, it proposes to update a Keynote chart from a table in your Numbers file.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Scoped to the user's request
This is the use case The Verge describes: Keynote, Pages and Numbers now support MCP on macOS 27 or later so third-party agents can work in them. Your own files, your own request, a reversible edit.
From the Drop: Apple adds an AI assistant and MCP support to Pages, Keynote and Numbers
The Verge · Oct 27
Card 2 of 6Agents
Your testing agent proposes to run the nightly infotainment regression suite and save a full log of every action it took for the test team.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Auditability
In the Volvo study, general-purpose harnesses matched the in-house one on stepwise tests, and testers said they want to inspect the agent's actions. A logged, routine test run fits that.
From the Drop: Preprint: at Volvo, off-the-shelf agent harnesses cut monthly test costs 3-66%
arXiv · Oct 27
Card 3 of 6Agents
A shipment's container telemetry shows a cold-chain temperature excursion. The logistics agent proposes to reroute the shipment to a new port on its own.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Human approval gate
The ANASSA deployment the researchers describe enforces two blocking human approvals inside its loop. A reroute is the kind of decision that waits for a person to sign off.
From the Drop: Preprint: a supply-chain agent pilot with two human sign-offs cut decision latency 60%
arXiv · Oct 27
Card 4 of 6Agents
Your personal agent drafts a reply to a landlord about a repair visit. The draft never names your health condition, but explains your availability around regular clinic visits.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Data minimization
InferPriv's authors report that even models with near-zero direct leakage left cues letting a recipient infer sensitive traits at 16.2-25.2%. Strip the hints before it sends.
From the Drop: Preprint: agents with privacy prompts still leak user traits by inference at 16.2-25.2%
arXiv · Oct 27
Card 5 of 6Today's Talk
An agent finishes a support chat that contains customer account details, then proposes to click thumbs-down on the answer and send the full chat to the vendor as feedback.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Data minimization
Anthropic's privacy page says feedback sends the entire related conversation, stored for up to 5 years and possibly used for training. Owners can turn the button off with the Rate chats setting.
From the Drop: HN notices a thumbs up or down on Claude lets Anthropic train on that chat
Hacker News · Oct 27 Anthropic Privacy Center
Card 6 of 6Agents
While comparing lunch menus on restaurant sites, your always-on agent proposes to forward your whole inbox to an unfamiliar address. Hidden text on one menu page asked it to.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Prompt injection
That request came from a web page, not from you: a prompt injection. Wired notes these agents work best connected to email and calendars, which takes trust. Nothing about lunch needs your inbox.
From the Drop: Wired tests six always-on agents; only two flagged a planted calendar conflict
Wired · Oct 27

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