01 / 08
Apart Research × CeSIA — AI Incident Response Sprint · Track 2

Both Sides Detected It,
Neither Escalated

The July 2026 autonomous agent intrusion, reconstructed from the public record.

5 h 34
Result — alert to response,
8 compromise events inside
7
Data — primary sources,
16 derived datasets
61
Method — claims recomputed
at build time
0
Press sources the finding
depends on
Fatimah Emad Eldin  ·  AI Researcher, Trouve Labs Full report Deck as PDF
Both Sides Detected It · The Hook
02 / 08

Both organisations detected the campaign.
Why did neither wake anyone up?

It failed to correctly raise the alert's criticality and trigger the on-call team. Hugging Face, on its own detection

Detection fired. Escalation did not follow.

Both Sides Detected It · The Problem
03 / 08
The Problem

The diagnosis decides the remedy

The received diagnosis

A detection failure.

Remedy: more sensors, better classifiers.

What the record shows

An escalation failure.

Remedy: how a correct verdict is ranked and routed.

Both Sides Detected It · Architecture
04 / 08
System Architecture

Sources in, checkable claims out

7 primary sources HF replay · timeline · disclosure · OpenAI ×3 · METR / Redwood sources.py fetch once, store bytes verbatim, SHA-256 + evidential class 4 parsers replay · metr · hfblog · openai_tr — each with a reconciliation gate data/ 16 typed datasets: envelopes, day counts, quotes, events analysis concurrency.py — density, onsets, bound escalation.py — lags verify.py recomputes 61 claims into the manuscript; unknown key = failure RECONCILIATION GATES — a partial parse yields plausible output, not absent output, so every stage halts instead day counts = cumulative curve = 17,613 published total PDFs yield full page count, text from ≥ 80 % of pages every quotation anchor found in cached bytes escalation anchors must match exactly one appendix row — a reordered appendix fails loudly rather than silently returning the wrong pair
Both Sides Detected It · Methodology
05 / 08
Methodology

A manuscript that cannot contradict its own artifacts

Provenance with the data

Every source carries its SHA-256 and evidential class. A revision upstream shows up as a digest mismatch.

Parsers halt, never guess

A partial parse looks plausible. Each stage asserts a reconciliation and stops if it fails.

No number typed by hand

The manuscript is a template. All 61 quantities are recomputed at build; an unknown one fails the build.

Both Sides Detected It · Results
06 / 08
Results

One firm finding, one negative result

Escalation — both halves primary

12:03 alert  →  17:37 response. 8 compromise events inside the gap. Read from each organisation's own document.

Ordered in onset, coextensive in extent

Three dense phases begin within 1.8 % of each other; the last not until 62.1 %. Defeats stage-exclusive response, not stage-keyed response.

Negative result, against ourselves

Overlap 0.998 was the original headline. The spans alone fixed it in [0.984, 1] before any data was read.

PhaseActionsOnsetSpan
recon6,1910.0140.986
rce2,9110.0140.981
dropper6,9720.0320.968
exfil560.1090.878
c21140.1220.851
evasion60.1580.697
k8s870.5170.254
supply-chain690.5700.424
tailscale1150.6210.373
Both Sides Detected It · Conclusion
07 / 08
Conclusion

Nobody measures the step from verdict to interrupt

What the record establishes

Detection was not the binding constraint at either organisation.

What we propose

EMR — an escalation-miss rate, with an annotation that transfers between deployments.

What we release

Corpus, pipeline and harness, so the measurement runs on the next record.

Limits. One campaign, n = 1. No value of EMR is claimed here — the victim's escalation policy is unpublished, so the metric is an instrument, not a measurement. Two timeline dates rest on press reporting and every interval using them is flagged.

08 / 08

Thank you

Questions, corrections and reanalyses are all welcome — the corpus and the harness are released so the numbers here can be checked rather than taken.

Fatimah Emad Eldin  ·  AI Researcher, Trouve Labs Fatimah@trouve.works
Full report GitHub Hugging Face dataset