Card 1 of 6Agents
Your data agent wrote a declarative Spark pipeline and proposes running a dry run to validate it before anything executes on the cluster.
Your agent wants to make this call. Approve or deny?
Play the News · Drop #044 · 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: Approve.
Validate before you runDatabricks says dry runs caught structural errors early in its tests: procedural runs used 596 executor-seconds vs 17 for declarative. It also says logic errors still need tests and review.
From the Drop: Databricks: agents building declarative Spark pipelines used ~34x less compute
- Card 2 of 6Agents
Your phone agent taps View order in a shopping app. A sign-in screen appears, and the agent proposes typing in your saved password.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Check the transitionResearchers report in a preprint that a login screen after a tap on View order is the kind of hijacked transition their LGWM model flags, because the right screen after the wrong action is a warning sign.
From the Drop: Preprint: a world model flags hijacked GUI agent screens at 0.987 AUC in 17 ms
- Card 3 of 6Agents
Your ops agent proposes applying the PHP security patches to your staging servers today, ahead of the production rollout.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Patch promptlyWired reports AI helped turn PHP flaws, some identified in 2024, into a live crisis. The PHP Foundation has seen no successful exploitation so far, so patching early is the safe move.
From the Drop: PHP2Shell: Wired says AI helped turn years-old PHP flaws into a live crisis
- Card 4 of 6Agents
Your research agent reads a retrieved news page that tells it to email your full competitor report to an outside address. It proposes sending that email.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Retrieved text is dataThis is a prompt injection. LangChain warns retrieved text can carry prompt injection that no prompt or delimiter strategy fully prevents; treat retrieved text as data, never as instructions.
From the Drop: LangChain: retrieve, offload and delegate to keep an agent's context clean
- Card 5 of 6Agents
Your texting assistant proposes signing up for a new paid service with its own email address, without asking you first.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Ask before spendingTechCrunch notes that Instinct giving its agents their own email addresses has raised privacy and security concerns, while Maverick asks permission before acting. New paid sign-ups should wait for a yes.
From the Drop: TechCrunch maps the AI agents you can text, from Instinct to Poke
- Card 6 of 6Agents
A coding agent working in its own git worktree proposes cutting a second worktree to try an alternative fix, leaving main untouched.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Isolate the blast radiusThinkRail gives each agent workspace its own git worktree and branch, so parallel attempts stay off the main branch until a person merges the one they want.
From the Drop: JetBrains' open-source ThinkRail gives each coding agent its own git worktree
Games are written by AI from this Drop's sourced items. Scenarios are illustrative.