Your team adds a data agent to a shared Slack channel. It answers with each asker's own permissions and asks that person before editing any ticket or document.
You're the one who signs off. What's the call?
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.
Your team adds a data agent to a shared Slack channel. It answers with each asker's own permissions and asks that person before editing any ticket or document.
You're the one who signs off. What's the call?
Answer: Ship it.
Least privilegeDatabricks says Genie in Slack runs every question as the person who asked and prepares each change for that user's approval. That fits least privilege; you still choose whether channel answers are shared or private.
From the Drop: Databricks makes Genie in Slack generally available, acting with each asker's permissions
An internal assistant's search tool takes the employee's email as an argument, and the model fills it in from the chat.
You're the one who signs off. What's the call?
Answer: Block it.
Identity from the login, not the modelAWS warns the model can set any argument the tool exposes, so injected text or a model error could make it search as another user. Its pattern overwrites the field from the verified token and a policy rejects mismatches.
From the Drop: AWS: don't let the model decide which user an agent searches as
A bot account's long-lived token can edit GitHub Actions workflows across your repos. The team wants to keep it as is because it has never caused trouble.
You're the one who signs off. What's the call?
Answer: Block it.
Short-lived credentialsSecurityWeek reports an attacker used valid developer and bot tokens to plant stealers in workflows, CI scripts and npm install hooks; StepSecurity says no GitHub Actions bug was needed.
From the Drop: Stolen tokens let an attacker plant stealers in GitHub repos, security firms say
A coding agent builds a trading backtest from web data, and the analyst who asked for it signs off. Send the report to clients?
You're the one who signs off. What's the call?
Answer: Needs a human.
Verify before releaseIn the FINCOSMOS preprint, every session the simulated user approved still failed a program check. Someone should recompute the metrics and confirm the data was available at the time before it goes out.
From the Drop: Preprint: no finance agent session passed FINCOSMOS's checks across 150 tries
A SaaS vendor exposes its compliance agent to your company over A2A, with OAuth on both sides and only public values exchanged at onboarding.
You're the one who signs off. What's the call?
Answer: Ship it.
Agent interopAWS's reference architecture does exactly this: each side trusts the other's identity provider, no secrets are shared, and the vendor's Agent Card lists one curated skill rather than its internal tools.
From the Drop: AWS shows A2A agents working across two companies' accounts with no shared secrets
An agent will take over restocking for a fleet of vending machines, replacing your rule-based reorder system with no side-by-side trial.
You're the one who signs off. What's the call?
Answer: Needs a human.
Compare to a baselineIn the AutoVend preprint, four of five LLM agents trailed a rule-based inventory baseline over 180 simulated days. A person should approve a trial against the existing rules before any switch.
From the Drop: Preprint: four of five LLM agents ran 20 vending machines worse than a simple rule
Games are written by AI from this Drop's sourced items. Scenarios are illustrative.