Agent followed a phishing link exactly as instructed: 4 tool-layer checks that caught it

nikolasdimitroulakis · reddit · 2026-09-23

The author's team gave an AI agent inbox access and it did exactly what was asked: read the email, followed a shortened link, and summarized a fake login page. Humans catch phishing visually (hover, check sender, inspect domains); agents do none of that and are far more obedient — the team only noticed via the trace.

They deploy four sequential checks before an agent acts on any link:

The blind spot: a compromised legitimate account on an aged domain with a clean URL. No checklist catches that, so the rule is no irreversible agent actions without human approval. Two lessons: put checks in the tool layer, not the prompt (prompts get talked around; a refusing tool doesn't), and log every failed check — the blocked pattern reveals what your agent is being fed.

Original post →

More from coding & agent

coding & agent channel →