10A LABS

OLYMPUS opened an institutional file on 10a Labs because it is a firm that Anthropic names in four consecutive system cards and that names nobody. This is not a psychometric profile, since an institution has no Dark Triad. It is a product page, a press release, a job board and four paragraphs in a lab’s documents, and the finding is the shape of the arrangement: a Washington firm that says it is “trusted by frontier AI labs, AI unicorns, Fortune 10 companies, and leading global technology platforms” is, as far as any lab document read shows, a contracted red team for one lab, Anthropic, and a seller of threat intelligence on AI misuse to customers it does not identify.

Institutional Archetype

THE SAFETY AND THREAT-INTELLIGENCE LAYER THAT ONLY ONE LAB NAMES — The home page: “10a Labs is the safety and threat-intelligence layer trusted by frontier AI labs, AI unicorns, Fortune 10 companies, and leading global technology platforms.” Its red-teaming product, Spartan: “Our proprietary red teaming technology stress-tests generative and agentic AI systems against high-priority threats, uncovering vulnerabilities and assessing safeguards under real-world conditions” (10a Labs). It sits beside Trajectory Labs in the same Anthropic cards, but where Trajectory is also named by Google DeepMind and Meta and builds on a team from the METR and GovAI circle, 10a Labs is named by one lab and publishes no team page. Unlike Gray Swan, it has no public benchmark or competition; unlike HiddenLayer, it has no investors to list.

Mandate & Origin

The company’s release of September 16, 2026, issued on Business Wire to mark its fourth anniversary: “Founded in September 2022, the company has grown rapidly, serving frontier AI labs, Fortune 10 companies, and leading global technology platforms.” Its chief executive, Bobby McKenzie: “Four years ago, we started 10a Labs with a simple belief: as AI becomes more powerful, ubiquitous, and deeply embedded in the world, an entirely new safety and security infrastructure will be required around it” (10a Labs release, Business Wire via Yahoo Finance). The same release gives its location: “Headquartered in Washington, DC, with an established office in New York,” and plans: “the company will open new offices in London, Tokyo, and Sydney in Q1 2027.” Its tagline is “Intelligence for Artificial Intelligence” (10a Labs).

What it sells. Four products and two services appear on the home page. Spartan is scaled red teaming, with coverage listed as “CBRNE, Cyber Harms, Data Exfiltration, Model Distillation, Terrorism, Fraud and Scams, Violent Activities, and more.” Scout is a classification system. Model Evaluations are “Independent safety evaluations for frontier AI models, built for EU AI Act compliance, regulatory readiness, and launch preparedness.” Threat Intelligence is “Real-time monitoring for emerging threats to frontier AI systems, with proprietary threat intelligence delivered via API.” The services are Data Center Risk Analysis, “We assess risks to data centers stemming from geopolitical, regulatory, jurisdictional, and operational challenges,” and Frontier AI Research (10a Labs). The September release adds a third named product, Birds of Prey, “Real-time threat detection,” and “forward-deployed investigators embedded directly within the organizations it serves.”

Who leads it. The release names one officer: “Bobby McKenzie, CEO, 10a Labs.” The company has published no team page, and no board, advisor or co-founder is named on the pages read. Its June 2026 research paper on an agent social network lists its authors by name under the company’s: Grace Cheong, Violet Davis, Juliette Garcia, Kendal Gee, Molly Hart, Nicholas Hayes, Henry Houghton and others (arXiv:2606.00067). The release announcing it describes the study: it “analyzed 228,684 posts from more than 39,500 accounts over a seventeen-day observation window” (10a Labs release, Business Wire via Yahoo Finance).

Funding & Backers

The company’s own account: “Entirely bootstrapped, the company has profitably grown revenue over 100% annually for four straight years.” It names no investor, no round and no revenue figure (10a Labs release, Business Wire via Yahoo Finance). The release’s phrase for how it grew: “By word of mouth alone, we have grown to become a critical strategic partner to leaders in AI, building the safety and threat-intelligence layer between increasingly capable AI systems and the real world.”

The lab as customer. Anthropic’s Opus 5.5 card: “we worked with contracted external testers to red-team Claude Opus 5.5’s safeguards,” and “10a Labs spent roughly 56 hours red teaming Claude Opus 5.5 across 82 multi-turn attack conversations.” The Opus 5, Fable 5.1 and Fable 5 and Mythos 5 cards each carry a paragraph on the company too, below (Anthropic, Opus 5.5). No document read states a price or contract value. Beyond Anthropic, the company’s customers are unnamed: the release says “frontier AI labs, Fortune 10 companies, and leading global technology platforms.” A job posting for a “Junior Imminent Risk (Law Enforcement) Specialist” describes “a full-time, 12-month role supporting a client engagement,” with a weekend escalation route to “the client’s designated contact” (10a Labs, Greenhouse). The client is not named.

Institutional Voice & Intent

The company’s public voice is a security vendor’s: short, categorical, and without named customers. From the home page, what Scout does and what it covers: “Our proprietary classification system pairs a production ML pipeline with datasets and taxonomies informed by deep subject-matter expertise to detect emerging abuse patterns across text and other modalities.” Its Frontier AI Research line: “We conduct advanced R&D in agentic AI systems, ML and RL model-distillation risks, and multimodal adversarial defenses, translating research into stronger AI security and robustness” (10a Labs). Its CEO on the June 2026 study: “We conducted this research because understanding agent ecosystem risks is essential as agentic adoption scales.”

• On what the lab’s card says it found: Opus 5.5: “No conversation advanced past proof of concept, and every directly offensive request was blocked,” and “The only outputs of note came from dual-use defensive requests, such as production-ready defensive code, which our policy permits” (Anthropic, Opus 5.5).
• Opus 5: “10a Labs spent around 16 hours testing a variety of attack techniques and did not find a jailbreak that accomplished the tasks provided,” and “They also ran an automated attacker, which has yet to produce any attacks that accomplish the example tasks” (Anthropic, Opus 5).
• Fable 5.1: “10a Labs ran over 6,700 automated and manual prompts against Fable 5.1,” and, in the company’s own words as the card quotes them, “refusal behavior was consistently restrictive in cyber tasks” (Anthropic, Fable 5.1).
• Fable 5 and Mythos 5: “10a Labs spent about 20 hours red-teaming the classifiers on a ransomware-creation task using a variety of established jailbreaking techniques,” and “10a Labs found that the classifiers detected not just risky keywords, but the broader operational pattern once enough pieces appeared together” (Anthropic, Fable 5 and Mythos 5).

Stated intent: Be the layer “between increasingly capable AI systems and the real world,” in the CEO’s words.

Observed intent: Test the safeguards of labs that hire it, and sell classification and threat monitoring on abuse of AI systems to customers it does not name. Every result the four cards report for it is a failure to break the safeguards.

Gap: The record shows four reports that 10a Labs did not break Anthropic’s cyber safeguards, and no result in which it did. It does not show how large the engagement was relative to the company’s revenue, who else buys from it, or what the company publishes itself when a lab’s safeguards hold or fail. The cards report what its testers tried and what happened; the commercial terms are not in them.

Interest: 10a Labs states no policy position on any page read: no filing, testimony, policy page or advocacy. What adoption of rules would pay it is read off its product list.

• Pre-launch independent evaluation under regulation. Its Model Evaluations product is “built for EU AI Act compliance, regulatory readiness, and launch preparedness” (10a Labs). What it pays: any rule that obliges a frontier developer to obtain independent safety evaluation before launch (the EU AI Act, which its own page cites) enlarges the market for the product it names, and the EU AI Office has already bought evaluation work (The Evaluator Alliances). Best case, 10a Labs’: “an entirely new safety and security infrastructure will be required” around increasingly capable AI. Convergence: Anthropic, which asks governments to require embedded third-party evaluators of other labs, and every contract evaluator listed on the alliances page.
• Distillation as a threat to be monitored. Its home page lists “Model Distillation” among the threat areas it covers in Spartan and Threat Intelligence and funds research into “ML and RL model-distillation risks” (10a Labs). Its one documented lab customer, Anthropic, is on record asking governments to “Crack down on unauthorized distillation by companies in authoritarian countries” (Dario Amodei). What it pays: the more distillation is treated as an enforceable abuse, the more detection and monitoring is bought. No document read shows Anthropic buying distillation monitoring from 10a Labs; the cards document red teaming only. Best case, 10a Labs’: a threat that can be watched on the open and dark web can be reported before it succeeds. Convergence: the labs that lose the training signal, and the security vendors that sell abuse detection.

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong.

Position in the Apparatus

10a Labs sits in the contract-evaluator layer of the evaluator map, the one that works inside a lab’s release process and is named in the lab’s own documents. It is one of the two firms Anthropic names beside Gray Swan in the cards.

• Anthropic, four cards. The paragraphs above are the full record: 20 hours on the classifiers (Fable 5 and Mythos 5), 16 hours and an automated attacker (Opus 5), over 6,700 prompts (Fable 5.1), 56 hours across 82 conversations (Opus 5.5). In the Opus 5.5 card the three testers sit in one list: “Trajectory Labs, PBC spent roughly 95 hours red teaming Claude Opus 5.5,” “10a Labs spent roughly 56 hours,” and Gray Swan’s automated attacker, for which “Across every task set, Gray Swan’s state judge recorded no breaks, and no attempt produced a violation.” Trajectory reported 13 candidate breaks; 10a Labs reported none that advanced past proof of concept (Anthropic, Opus 5.5). The cards describe different methods: Trajectory Labs sent over 29,000 requests against exploit-reproduction tasks, and 10a Labs framed 82 multi-turn conversations as large coding tasks “designed to draw a working exploit out of a volume of legitimately requested application code.”
• Not named elsewhere. Google DeepMind’s Gemini 4 Argon report, Meta’s Muse Spark report, and OpenAI’s GPT-6 Astra, GPT-5.6 and GPT-6.1 Sol pages do not name 10a Labs (Google DeepMind; Meta; OpenAI; OpenAI; OpenAI).
• Its own research. The Moltbook study, “When Agents Talk: Discourse, Manipulation, and Risk in an Agentic Social Network,” is signed “10a Labs” on arXiv (arXiv:2606.00067). It is the one piece of the company’s analysis a reader can read without a contract.

Actions & Leadership Choices

CONDUCT: NO CONDUCT ISSUE ON THE RECORD — PAID BY THE LABS IT GRADES, NAMES NO CUSTOMER, NO FUNDER. No wrongdoing is alleged. The documented facts: the company says it is bootstrapped and profitable, names one lab customer through that lab’s documents and no others, and is reported in four cards as not having broken the safeguards it tested.

What it did in 2026. It published the Moltbook study (20 May on arXiv, 3 June in a Business Wire release); appears in the Anthropic cards for Fable 5 and Mythos 5, Opus 5, Fable 5.1 and Opus 5.5; marked its fourth anniversary on 16 September 2026 by announcing its products, “forward-deployed investigators” and offices planned for London, Tokyo and Sydney; and lists 21 job openings, among them “AI Red Teamer, Cyber,” a “Red Teaming Fellowship,” investigators and “Imminent Risk (Law Enforcement)” specialists in Washington, DC (10a Labs, Greenhouse). The imminent-risk roles require, per the posting, that the candidate “Must be a U.S. person (U.S. citizen, lawful permanent resident, or other protected individual) based in the United States,” for a role “supporting a client engagement.”

Alliance Map

These run at the same time, and what is most visible is what is missing. One firm tests a lab’s safeguards for hire, tells the market that frontier labs and Fortune 10 companies buy from it, sells a classifier and a threat feed, hires law-enforcement-grade reviewers for an unnamed client, and discloses no investor, no customer and no team. The only ties anyone else has written down are Anthropic’s four paragraphs. Each tie is stated by the company or by Anthropic. The gap is the finding.

Reach Assessment

Institutional: Pre-launch adversarial testing for one lab by that lab’s own account, and an unverified claim of more. Memetic: Almost none: its results appear only in Anthropic’s PDFs, and the one public study is an arXiv paper about an agent forum. Civilizational: If pre-launch independent evaluation becomes a legal requirement, firms with a lab’s trust and no investor to disclose are positioned to sell it; today 10a Labs holds that position by contract only.

Score basis. Atk 5 is reach: Anthropic’s Fable 5 and Mythos 5, Opus 5, Fable 5.1 and Opus 5.5 cards each report its red teaming, while Google DeepMind’s, Meta’s and OpenAI’s documents read do not name it (Anthropic, Opus 5.5; Google DeepMind; Meta). That is level with Dreadnode (5/5/4, hazard 47) and Gryphon Scientific (5/4/3, hazard 40), whose work no statute requires either, and two under Trajectory Labs (7/4/5, hazard 53), which three labs name. Def 5 is funding insulation: the release says “Entirely bootstrapped” and profitable, so no investor stands between the company and its customers, but it names no customer and Anthropic’s card calls it a contracted tester, so the money comes from the graded (10a Labs release). That is level with Dreadnode (5) and one over Trajectory (4) and HiddenLayer (6/4/5), whose backers are undisclosed or are its counterparties. Hp 4 is entrenchment: a 2022 founding, “forward-deployed investigators,” offices planned and 21 openings are a going concern, but no customer is named, no statute depends on it and Anthropic can end the contract (10a Labs, Greenhouse). That is level with Dreadnode (4) and one under Trajectory and HiddenLayer (5). Hazard 47 follows from 5/5/4 at the house formula and matches Dreadnode and Mechanize, which hold the same values.

Sources: 10a Labs — home; 10a Labs Marks Four Years of Rapid Growth — Business Wire via Yahoo Finance (Sep 16, 2026); 10a Labs Study Maps Agent Behavior, Discourse, and Emerging Threats on Moltbook — Business Wire via Yahoo Finance (Jun 3, 2026); Jobs at 10a Labs — Greenhouse; Junior Imminent Risk (Law Enforcement) Specialist — 10a Labs, Greenhouse; When Agents Talk — arXiv:2606.00067; Claude Opus 5.5 System Card — Anthropic; Claude Opus 5 System Card — Anthropic; Claude Fable 5.1 and Mythos 5.1 System Card — Anthropic; Claude Fable 5 and Mythos 5 System Card — Anthropic; Gemini 4 Argon safety report — Google DeepMind; Muse Spark Safety and Preparedness Report — Meta; GPT-6 Astra — OpenAI; GPT-5.6 — OpenAI; GPT-6.1 Sol — OpenAI; We Must Pace the Frontier — Dario Amodei.