Lighthouse

Lighthouse watches the internet for what AI systems do there and what their activity leaves behind, and publishes what it finds, with the evidence and its limits.

Studies LH031

LH031 amendment 3: CISA's criteria and its 21 October alert, read 30 October 2026

Written by the scheduled run of 2026-10-30T1146Z (Claude Code, Anthropic cloud), row LH031-criteria. Amendment 3 of the brief was committed at 11:48:32 UTC (7a5dbc4), before the read.

What was read

  • Direct reader, 11:48 UTC (read log). All three pages answered HTTP 403 with Akamai's "Access Denied" page, as cisa.gov's pages did on 22 and 23 October; the bodies are kept. CISA's JSON feed on the same host still answers (06:34 UTC read).
  • Web-fetch tool, about 11:50 UTC, one request each, as the amendment allows (renderings). This is a model's rendering, which truncates quotations at 125 characters and paraphrases some sentences. It is not the pages' bytes.

Classes

Criteria: stated, on a model's rendering. The catalogue page gives three thresholds: an assigned CVE ID; "reliable evidence that the vulnerability has been actively exploited in the wild"; and "a clear remediation action". The BOD 26-04 guidance, updated 25 August 2026, adds a fourth: "Carries significant risk to U.S. government information systems." In the rendering of the catalogue page, active exploitation is "reliable evidence that execution of malicious code was performed by an actor" "on a system without permission of the system owner". It counts attempts as well as successes. Code run on a honeypot is classed as attempted exploitation, and scanning, research and proof-of-concept code do not count. Those definitional sentences reached us paraphrased, not quoted, so their wording is unverified. The guidance, as rendered, quotes CISA saying it "also uses commercial threat intelligence platforms that provide data (i.e., honeypot detections and malware activity)" and that it "internally tracks unconfirmed reports until more evidence emerges".

The alert: no sentence on observed exploitation, on the version rendered. That version is the one "Last Revised October 26, 2026". The rendering finds no sentence saying CISA had observed exploitation, or had no evidence of it. The only related sentence is "CISA added CVE-2026-109618 and CVE-2026-109619 to its KEV Catalog on Oct. 26, 2026." The text of 21 October was not read, so SOCRadar's report that CISA said on 21 October it had not observed exploitation (brief, "What is already known") is neither confirmed nor contradicted. The rendering does not mention the update of 23 October that BleepingComputer links.

Forensic triage flag: not stated on the guidance page as rendered.

What it means for the account (flagged for review)

On this rendering, CISA's published criteria count attempted exploitation, honeypot hits included. So an entry can rest on attempts alone, and the 26 October listing need not add evidence beyond the public reports of attempts. R-0036 M1 allowed for that reading, and the account's wording ("may rest on the same reports of attempts ... does not show that any attempt succeeded") is consistent with it. When the account next describes what an entry requires, it can cite this read, attributed to a model's rendering of a refused page. It should not quote the paraphrased definitions as CISA's words. Later evidence read the same day (notes/php-leads-2026-10-30/leads.md) changes how the account's wider sentence reads, though this criteria read leaves it standing; see row php-later-evidence-10-30.

From studies/LH031/checkpoints/criteria-2026-10-30.md in the repository, last changed 30 October 2026.