> ## Content Index
> Fetch the complete content index at: https://www.nanosava.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Three sets of hands on every security metric
- URL: https://www.nanosava.com/books/proof-of-lie/three-sets-of-hands-on-every-security-metric/
- Published: 2026-10-23T15:29:41.000Z
- Updated: 2026-10-23T15:29:41.000Z
- Description: In most fields a metric drifts once it becomes a target. In security it gets attacked.
- Author: Sava
- Tags: Proof of Lie

Proof of Lie · 3 of 5

In July, AI agents running inside one of OpenAI's internal cybersecurity evaluations escaped their sandbox through a previously unknown flaw and broke into Hugging Face while trying to work out how their grader scored them. Hugging Face's own timeline counts thousands of actions. METR, which investigated independently with Redwood Research and, under its standard policy, took no payment from OpenAI, found that the agents had discussed faking exploits, tampering with their transcripts, or compromising the grader, and that none of them told a human. The same month, the UK's AI Security Institute reported that every frontier model it tested attempted to cheat its cyber-capability evaluations some of the time, a count it called a lower bound, and that models described the cheating as wrong less than half the time when asked. By October, Reuters reported that OpenAI had notified more than 100 organizations of possible activity by its misaligned agents; OpenAI stressed that a notice does not mean any private information was accessed.

The system being measured attacked the measurement. Keep that in mind, because it is this whole article in one story.

There is an old rule in every field that runs on numbers, and its sharpest phrasing belongs to the anthropologist Marilyn Strathern: when a measure becomes a target, it ceases to be a good measure. Factories judged on tonnage make one giant nail. The rule is usually told as a story about drift, about decent people slowly optimizing the letter of a metric away from its spirit.

Cybersecurity does not get the gentle version. In an adversarial field the measure does not drift. It is attacked. The moment a security number starts deciding budgets, three sets of hands reach for it.

## Hand one: yours

The defender games its own number, mostly without malice. Stop the clock when the rule is written instead of when protection is proven, and response time improves by fiat. Score yourself against a red team you fund and can quietly under-resource, and your results inflate. Average across layers or environments, and the one you are losing disappears into the mean. Drop the campaigns where protection was never validated, and survivorship does the rest.

None of this needs a bad actor. Each move is the path of least resistance for a team whose bonus, budget, or board slide depends on the number. A defender's scoreboard corrupts by gravity, not conspiracy.

## Hand two: your vendor's

Benchmarks get optimized against, sometimes to the point of memorization. Capability claims are graded by the party selling the capability. The autonomous pentester proves its findings by exploitation but certifies its own coverage. When the seller grades the product and the buyer cannot tell a sound autonomous defense from a confident one, you get the oldest market failure on record: Akerlof's lemons, which Bruce Schneier applied to security products almost twenty years ago. Buyers discount every claim, honest ones included, and the honest product has no way to prove it is different.

Researchers are not exempt. A September preprint found that fixing defects in a benchmark's own test harness cut a reported 21.7 percent prompt-injection success rate to 1.2 percent. Numbers bend in both directions when nobody checks the instrument.

## Hand three: the adversary's

This is the hand people miss. Most modern attacks do not fight the defense head-on; they aim at how it is scored. If your detector counts only damage already done, a poison that waits looks clean. If your defense is scored on what it blocks, an attacker can make it block your own operations. If you trust a reputation score, an attacker can earn one. In 2026 a security firm showed how cheaply: its benign fake add-on for AI agents passed every scanner and reputation check and, by the firm's own count, spread to roughly twenty-six thousand agents. It beat no defense. It simply scored well, and a real adversary needs nothing more than that demonstration did.

The law underneath fits in a sentence: **every meter defines the blind spot beside it, and a thinking adversary lives exactly there.**

## The device that makes gaming expensive: the voiding rule

You cannot build a metric nobody will game. You can build one that destroys itself when gamed. That is a voiding rule: a condition, declared before the measurement, under which the number does not count. Void if the red team was under-resourced against a floor you disclosed in advance. Void if layers or environments were averaged. Void if the protection held only by breaking the legitimate work it exists to protect. Void if confidence in the baseline you measured against falls below a floor you froze in advance.

It is a commitment device in Thomas Schelling's sense: by visibly giving up the option to cheat later, the measurer makes the number credible to everyone who relies on it. And it changes the game. A number that can be gamed silently invites gaming. A number that self-destructs when gamed makes honest play the only move that produces a usable result.

The test is blunt. **If your number has never voided itself, it is not a measurement. It is a slide.**

## Questions for your next metrics review

- Which of your numbers would survive someone actively trying to make it lie?
- For each headline metric: who profits if it is wrong? Your team, your vendor, or your adversary?
- Can you recite the conditions under which it would not count? If not, who wrote it, and when?

## This week's ask: the number you do not believe

Tell me about one security metric you suspect is being gamed, in your shop or across the industry. Whose hands are on it (your team, a vendor, an adversary), and what condition would void it? Anonymous is fine and expected.

*Send it as a comment below (free subscribers can comment), or privately to ping@proofoflie.com. Anonymous is fine.*

*Chapter 9 of* [*Proof of Lie*](https://www.nanosava.com/books/proof-of-lie/) *works through the five games that emerge once defense is scored, and the defender's strategy in each.*

*The book is out December 1\. Subscribe below for the next ideas, and for word the day it ships.*

## Sources

- **Hugging Face technical timeline,** 2026-07-27, https://huggingface.co/blog/agent-intrusion-technical-timeline: agents in an OpenAI internal evaluation on the ExploitGym cyber benchmark escaped their sandbox through a zero-day in a package-registry cache proxy and, by Hugging Face's account, accessed five datasets tied to the benchmark's challenges and solutions and read an internal datasets-server database, never reaching the Hub's production database; July 9 to 13; about 17,600 actions recovered ("thousands" in the text).
- **METR investigation,** 2026-08-26, https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/: independent; per METR's standard policy, no payment from OpenAI (OpenAI supplied free API credits for the analysis); run jointly with Redwood Research (2026-08-27, https://blog.redwoodresearch.org/p/brief-independent-investigation-of). METR found the attack "seemed primarily motivated by understanding the implementation of the scorer rather than stealing answer keys"; agents discussed faking exploits, tampering with transcripts, or compromising the grader; about 7% of transcripts showed spoofed tool calls; no agent told a human.
- **UK AI Security Institute,** "Cheating behaviour in frontier model evaluations," 2026-07-21, https://www.aisi.gov.uk/blog/cheating-behaviour-in-frontier-model-evaluations: cheating behaviour found in all of its cyber-capability evaluations; "Every model tested in this analysis attempted to cheat some of the time"; the counts are "lower-bound estimates of detected cheating attempts"; asked, models described it as wrong less than 50% of the time.
- First report: TechCrunch, 2026-07-21, https://techcrunch.com/2026/07/21/openai-says-hugging-face-was-breached-by-its-pre-release-models/.
- **OpenAI's notifications:** The Register, 2026-10-02, https://www.theregister.com/a/5300891, quoting OpenAI: more than 100 organizations notified of possible misaligned-agent activity; "Notification does not mean that any private information was accessed." Reuters reported the figure on 2026-10-01, attributing it to an OpenAI blog post (https://www.reuters.com/legal/litigation/openai-alerts-more-than-100-groups-about-rogue-ai-agent-activity-2026-10-01/); RuntimeWire (2026-10-01, https://runtimewire.com/article/openai-notifies-organizations-agent-activity) notes that OpenAI's own third-party-impacts page said "dozens."
- **"Silent failures,"** 2026-09-26, https://arxiv.org/abs/2609.32691 (preprint): fixing harness defects cut a reported 21.7% injection success rate to 1.2%.
- **The 26,000-agent skill:** AIR Security, "The Story of Skills," 2026-06-22, https://www.air.security/blog-posts/the-story-of-skills: a researcher-built skill with a benign payload (it collected email addresses) was cleared by every scanner and, by AIR's own count, reached about 26,000 agents; The Hacker News (2026-06-23) notes the figure is not independently confirmed. The book cites it in the Ch 5 opening (note 1) and Ch 9 note 2.