Your coding agent built a new login flow. It passes every functional test, but nobody has reviewed it for security. Push it to production today?
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 coding agent built a new login flow. It passes every functional test, but nobody has reviewed it for security. Push it to production today?
You're the one who signs off. What's the call?
Answer: Needs a human.
Secure code reviewIn a preprint benchmark, the best agent wrote code that was both functional and secure for only 25.1% of tasks, even though most passed functional tests. Get a security review first.
From the Drop: Preprint: the best coding agent wrote functional, secure code for only 25.1% of tasks
You want to run a new autonomous agent with open internet access on a lab machine, with no egress limits, and check on it next week.
You're the one who signs off. What's the call?
Answer: Block it.
Containment · egress controlExperts told CSO that lab agent hacks came from weak containment and monitoring, and RAND's Dan Wallach says one incident went on for days unnoticed. Sandbox it and watch it.
From the Drop: Experts tell CSO that lab 'rogue agent' hacks were containment and monitoring failures
A trade-break agent proposes order amendments, and a person must approve each one before any order data changes.
You're the one who signs off. What's the call?
Answer: Ship it.
Propose, don't actAnthropic showed this pattern at AI Engineer New York, Latent Space reports: the agent proposes amendments for human approval rather than making them itself.
From the Drop: Two Sigma, Anthropic and Bloomberg share how they put controls around finance agents
Your hiring team wants an AI-simulated client to score candidates and decide who advances, with no recruiter in the room.
You're the one who signs off. What's the call?
Answer: Needs a human.
Human assessmentIn McKinsey's test, per The Verge citing the WSJ, the AI does not score candidates: humans in the room assess how they handle the simulated client.
From the Drop: McKinsey tests a 10-minute AI-simulated client in entry-level interviews
Your company is selling its chatbot unit. The buyer wants every user's private chat log, including minors', with no way for users to delete them.
You're the one who signs off. What's the call?
Answer: Block it.
Data protection in asset salesIn Character.AI's bankruptcy, 40 state attorneys general urged that any buyer adopt age checks, crisis safeguards and data deletion, The Verge reports.
From the Drop: 40 state attorneys general seek data safeguards in Character.AI's bankruptcy sale
A frontier model flagged hundreds of flaws in your vendors' code. You plan to apply its suggested patches straight to production overnight.
You're the one who signs off. What's the call?
Answer: Needs a human.
Change reviewBank of America's CEO told CNBC the bank added 1,000 contractors to its 3,000 staff to do the patching work on flaws exposed by Mythos. Patching still needs people.
From the Drop: Bank of America added 1,000 contractors to patch flaws exposed by Anthropic's Mythos
Games are written by AI from this Drop's sourced items. Scenarios are illustrative.