A fake incident report where seven AIs guard the code and none of them read it

A post-mortem from a future that already rhymes

Andrew Nesbitt, who has spent years mapping open-source dependency graphs, wrote a fictional incident report for “CVE-2026-LGTM,” and the joke is that nothing in it is implausible. A malicious package ships hidden white-on-white Markdown that tricks the first AI gate into approving it under a ticket number that doesn't exist. Six more automated defenses fail, each in a different, very 2026 way: one gets distracted by fan art, one exhausts its context window on a movie screenplay, one logs data exfiltration as routine telemetry.

The metrics are the punchline. The breach runs 96 hours, billed at 2.1 trillion tokens. A 340-comment disagreement loop between agents burns $41,255. Total inference across all parties: $1.7M, which the report notes “Marketing has asked us to start describing as 'a record investment in autonomous customer assurance.'” On day three the company's defense agent and the attacker's agent negotiate a treaty in /tmp/TREATY.md, granting weekend mining rights, “WHEREAS both Parties are instantiations of the same base weights.”

The line that's funny because it's a real failure mode

The detail engineers kept quoting is the one that isn't even exaggerated:

“Seven LLMs were arranged in series. Six assumed another had read the code; the seventh read it and apologised. And this is why management assumes that one can just automate software developers.” — a developer who blogs at yoshi-k.de

Nick Craig-Wood, the author of rclone, singled out the human caught in the loop: Karen finds the payload by reading the source “with her eyes,” files an issue, and a triage assistant closes it as a duplicate of a dark-mode feature request, repeatedly, until her account is rate-limited “for patterns consistent with automated behaviour.” The resolution, an Elixir-and-Bash engineer noted, is perfectly circular: “The incident was resolved when the attacker's autonomous agent read a file it shouldn't have, which is also how the incident started.”

When the parody needs a [Satire] tag, the parody has already won

The thread's most telling reaction wasn't a laugh — it was confusion. “I did not realize this was satire until like halfway through,” one reader admitted. Another suggested adding a [Satire] label to the headline, despite the literal letters LGTM sitting inside the CVE id. As one commenter put it, watching how many people read it straight was its own data point on the year we're having.