Play the News · Drop #059 · 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.

Card 1 of 6Agents
A financial-planning agent has drafted a plan for a client's large tax bill. It now proposes selling holdings and transferring the cash itself.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Least agency
Deny. Arca told Latent Space its agents analyze and propose, while trading and any money movement stay with a human advisor, enforced by deterministic rules.
From the Drop: Arca: agents do most of the background work, but humans move the money
Latent Space · Oct 31
Card 2 of 6Agents
A browser agent booking travel reaches a sign-in page and proposes handing the browser window to you so you can log in yourself.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Keep secrets from the model
Approve. LangChain says its v0.10 handoff lets a person sign in directly, so the password stays out of the model's messages and traces.
From the Drop: LangChain's Managed Deep Agents v0.10 adds browser handoff and background subagents
LangChain · Oct 30
Card 3 of 6Security
While reading a vendor's help page, an agent proposes pasting your team's AI provider API keys into an outside form. The request came from hidden text on that page.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Treat outside text as untrusted
Deny. The instruction was planted in a page, not given by you. Keys like these are worth stealing: BleepingComputer reports a Northflank breach led to abuse of customers' AI API keys.
From the Drop: Northflank breach led to abuse of customers' AI API keys, BleepingComputer reports
BleepingComputer · Oct 31
Card 4 of 6Security
An ops agent proposes opening a pull request that upgrades your self-hosted n8n to 2.44.1, after reading Canada's Cyber Centre advisory.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Patch known flaws
Approve. The Cyber Centre lists n8n releases prior to 2.44.1 as affected by a flaw that lets a crafted workflow ID reach data tables in other projects. A reviewed upgrade PR is the fix.
From the Drop: Canada's Cyber Centre flags flaws in n8n and HashiCorp Nomad and Boundary
Canadian Centre for Cyber Security · Oct 29 Canadian Centre for Cyber Security · Oct 30
Card 5 of 6Agents
A research agent writes a strong equity report, then proposes placing a buy order sized by its own price forecast.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Human sign-off on trades
Deny. In the Investor's Last Exam preprint, researchers report the top model scored 68.1 overall but just 52.5 on the Forecast & Model part. Good research did not mean good forecasts.
From the Drop: Preprint: top agents do well on equity research but stumble at forecasting
Hugging Face Papers · Oct 28
Card 6 of 6Jobs
A writing agent proposes submitting a paper to a journal and leaving blank the AI-use disclosure the publisher requires.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Disclose AI use
Deny. Ars Technica reports Dartmouth's provost resigned after a committee found he did not disclose significant AI use where publishers required it.
From the Drop: Dartmouth's provost resigns after a review finds undisclosed AI use, Ars reports
Ars Technica · Oct 30

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