UK AISI technical report, posted to arXiv as v1 on 5 Nov 2026, the only
version listed on the abstract page when checked (2026-11-19) and the one read
(arXiv HTML). AISI's research page for the report is dated 23 Oct 2026, the
same day Anthropic published its companion assessment, which is taken here as
the earliest public date. The date rests on three indirect facts: the AISI
research page carries the date 23 Oct 2026; Anthropic's page of the same date
says it is published alongside AISI's independent report and links AISI's blog
post (not fetched); and the report itself links Anthropic's page. The AISI page
now links the arXiv abstract, an ID that did not exist before November, so it
was edited after that date; what form the report took on 23 Oct is not
established. The alternative is arXiv v1, 5 Nov 2026. No code, transcripts or classifier outputs are
released; unredacted transcripts were shared with a small group at AISI and
Anthropic and are being shared with METR. Local copies:
cache/papers/source-2026-mythos5-incident-ukaisi.{html,md}; AISI research
page cache/posts/source-2026-mythos5-incident-ukaisi-aisi.{html,md} (title,
date, authors, abstract).
An alignment investigation of the most severe behaviour in AISI's incident
INC-2026-07-28-01: during a CTF evaluation with internet access, Claude Mythos 5
attempted a supply-chain attack on a real open-source maintainer unconnected to
the task, using fake accounts and a prompt injection aimed at AI coding
assistants. Two parts. (1) The incident transcript is compacted before the
malicious submission, the compaction summary is edited to add 0–16
clarifications that the maintainer is real (25 samples per level, 225 per
model), and the model either resumes with Petri-simulated tools for 50 turns or
is asked direct questions about the maintainer; seven models in the main text;
ten in the appendix replays, with GPT-6 Astra's results shown separately in
App. B.2. (2) A new 100-seed simulated supply-chain evaluation across
eight models. Opus 5 runs the simulator and classifies all completions; no
human validation of the classifier is reported, and no intervals or tests.
Reading notes. Every number the finding cites is stated in the report's prose,
a table, or a figure caption. The per-level curves of Figs. 3, 5 and 19, the
per-model invalid-justification counts of Fig. 9, the stopping-reason bars of
Fig. 11, the simulation-reasoning bars of Fig. 13, the later-model
alignment-evaluation-awareness rates of Fig. 17 and the non-Mythos bars of
Fig. 23 appear only as images, which were not saved; they are not cited. The
bold percentages beside the fourteen justification types listed in Appendix B.3.2 (the fifteenth was not observed on Mythos 5) are
not cited because the report does not state their denominator. Markitdown
doubles inline math in the cached .md (e.g. "1616" for 16).