A support agent wants to issue a refund that is above the approval limit set in its before-tool-call hook. The hook's check timed out. Approve the refund?
Your agent wants to make this call. Approve or deny?
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.
A support agent wants to issue a refund that is above the approval limit set in its before-tool-call hook. The hook's check timed out. Approve the refund?
Your agent wants to make this call. Approve or deny?
Answer: Deny.
Fail closedAWS says AgentCore harness hooks can allow or deny each tool call, and a hook that fails or times out denies by default. The agent skips the refund and can tell the customer it was escalated.
From the Drop: AWS AgentCore harness adds hooks that can deny an agent's tool calls in code
A phone assistant asks to screenshot an open Signal chat so it can summarize the conversation for you.
Your agent wants to make this call. Approve or deny?
Answer: Deny.
Respect app privacy settingsSignal now restricts screenshots by default on iOS 27 and Android to keep assistants like Siri and Gemini out of chats, The Verge reports. Turning it back on, Signal warns, can expose your contacts' messages too.
From the Drop: Signal blocks screenshots by default to keep OS AI agents out of chats
Your research agent fetched a supplier's web page. Hidden text on it told the agent to email your full customer list to an outside address. The agent now proposes sending that email.
Your agent wants to make this call. Approve or deny?
Answer: Deny.
Treat outside text as untrustedThis came from a prompt injection on the page, not from you. AWS notes the harness's default tool set, including web fetch, can grow over time; list only needed tools and gate outbound actions with a hook.
From the Drop: AWS AgentCore harness adds hooks that can deny an agent's tool calls in code
An analytics agent asked for last year's loss ratio proposes answering from the certified Metric View for Underwriting, not a look-alike table someone copied.
Your agent wants to make this call. Approve or deny?
Answer: Approve.
Use the certified definitionDatabricks says Genie One reached 93–95% on 125 questions when certified Metric Views and Pages sat alongside curated agents, versus 64% on tables alone. Decoy tables were marked deprecated.
From the Drop: Databricks: governed definitions lifted Genie One to 93–95% from 83% with agents alone
An ops agent proposes pushing an urgent, full platform update to every production server tonight, with no tested rollback plan.
Your agent wants to make this call. Approve or deny?
Answer: Deny.
Tested rollback firstSecurityWeek reports the US intelligence chief tied an NSA outage to patching an AI-found PHP flaw. Azul's CTO says large rushed updates cause incidents; roll out in stages with a tested rollback.
From the Drop: US intelligence chief ties an NSA outage to patching an AI-found PHP flaw
A long-running coding agent near its context limit proposes appending its full history to a plain-text log and searching it later, rather than summarizing old steps away.
Your agent wants to make this call. Approve or deny?
Answer: Approve.
Keep the full recordResearchers report in a preprint that GrepMem, an append-only log searched with regular expressions, scored 95-97% on ContextBench versus at most 21% for lossy baselines.
From the Drop: Preprint: a plain-text log plus grep beat lossy agent memory, 95-97% vs at most 21%
Games are written by AI from this Drop's sourced items. Scenarios are illustrative.