HANDSHAKE AI

OLYMPUS opened an institutional file on Handshake AI because it is the one name in Meta’s Muse Spark report that belongs to a jobs marketplace: the report credits “a third-party human red team from HandshakeAI” with “approximately 149 hours” of attack time against Meta’s model before release (Meta, Muse Spark Safety and Preparedness Report, 26 May 2026). This is not a psychometric profile, since an institution has no Dark Triad. It is a mandate, a board, a Form D trail, a court docket and a voice, and the finding is the shape of the arrangement: a college-recruiting network that reports 25 million members now says it sells expert labor to eight labs, and some of that labor is paid to break the model of one of them.

Evaluator record: the engagements, the terms each document states, the findings quoted and what followed.

Institutional Archetype

THE CAREER NETWORK WITH A JAILBREAKER BENCH — Handshake describes itself on its about page as “the career network for the AI economy,” a “three-sided job marketplace” and says “Handshake AI” is “a human data labeling business that leverages the scale of the largest early career network” (Handshake, About; Handshake job posting, Notable Capital job board). Meta’s report calls the arm a “third-party human red team.” Unlike Gray Swan, whose automated attacker is the product, and unlike Trajectory Labs, which Meta’s report names beside it for a prompt-injection benchmark, the Handshake bench is people recruited through a marketplace: its posting for the role asks for “Creative, adversarial problem-solving skills” and says “No technical background is required” (Handshake job posting).

Mandate & Origin

Handshake was founded in 2014. The chief executive’s letter: “In 2014, we founded Handshake to change how, where, and why the next generation of talent builds their career—starting from their first day of college” (Handshake, About). The about page lists the co-founders as Garrett Lord (co-founder and CEO), Ben Christensen and Scott Ringwelski.

The AI business. Handshake’s own account of when it began: “In 2025 we introduced Handshake AI on a simple idea: that the expertise in our network could help shape how AI models learn. Today Handshake AI works with leading AI labs and, increasingly, with enterprises” (Handshake, Welcoming Nishant Ranka, Wes Field, and Stephen Rocco Rodi). TIME’s profile of Lord, in TIME100 AI 2026, dates the turn: “Last October, Handshake CEO Garrett Lord announced a strategy shift for the AI era, turning a careers platform for students into one of the fastest-growing suppliers of human expertise to AI developers” (TIME, TIME100 AI 2026: Garrett Lord, 27 Aug 2026).

Legal entity. The Handshake AI terms of service name the operator: “Stryder Corp. d/b/a Handshake” (Handshake AI, Terms of Service, updated 20 Aug 2026). The SEC’s EDGAR record lists Stryder Corp., CIK 0001666438, a Delaware corporation with its business address at 225 Bush Street, Floor 12, San Francisco (SEC EDGAR, Stryder Corp.). A federal docket adds a second defendant, “Handshake AI Solutions, LLC” (see Actions & Leadership Choices).

Who runs it. The about page lists the leaders: Garrett Lord, co-founder and CEO; Jonathan Stull, president and chief operating officer; Matt Greenberg, chief technical officer; Jesse Hulsing, chief financial officer; Dylan Lowrey, general counsel; Sahil Bhaiwala, chief strategy and innovation officer of Handshake AI; Wes Field, chief commercial officer of Handshake AI; Yang Zhao, chief business officer of Handshake AI; and Nishant Ranka, “Founder and President, Handshake Labs.” Its board: Lord, Ben Christensen, Margo Georgiadis (CEO, Montai Therapeutics), Mamoon Hamid (general partner, Kleiner Perkins), Alastair Mitchell (partner and co-head, EQT Ventures), Megan Quinn (former COO, Niantic), Toni Schneider (partner, True Ventures), Will Reed (general partner, Spark Capital) and Jeff Richards (managing partner, Notable Capital) (Handshake, About). Handshake’s announcement of Ranka says he spent six years building frontier models, “founding PM of Google’s MUM models and platform in 2021,” and “most recently worked on Microsoft’s MAI models” (Handshake, Welcoming Nishant Ranka, Wes Field, and Stephen Rocco Rodi).

Funding & Backers

The filings. Stryder Corp. filed five Forms D with the SEC between 2016 and 2021. The latest, signed 2 June 2021, reports a total offering amount of $79,999,697, all sold, to 12 investors, and lists as directors Garrett Lord, Ben Christensen, Toni Schneider, Mamoon Hamid, Megan Quinn, Alastair Mitchell and Margaret Georgiadis (SEC Form D, 2 Jun 2021). The Form D of 2 November 2020 reports $79,999,829 sold to 11 investors (SEC Form D, 2 Nov 2020); that of 30 October 2018 reports $39,999,936 sold to 14 investors (SEC Form D, 30 Oct 2018).

The valuation, in the company’s words. Its red-teaming job posting: “Capitalized & Scaling: $3.5B valuation from top investors including Kleiner Perkins, True Ventures, Notable Capital, and more” (Handshake job posting). The posting is hosted on the job board of one of the investors it names. No Form D after June 2021 is on the SEC record for the entity.

Revenue and contracts. TIME: “Since launching Handshake’s AI unit, Lord says it has grown from zero to more than $1 billion in gross annualized revenue in 16 months, working with eight frontier AI labs and drawing on hundreds of thousands of contributors” (TIME). Handshake’s enterprise page repeats the lab claim (“8/8 largest frontier labs train with Handshake AI datasets”) and says “70% of the Fortune 100 run their early talent operations on Handshake” (Handshake, Enterprise AI). The Handshake AI fellowship page reports “$500M+ payouts made” and “188k+ fellows” (Handshake AI). Meta’s report states no price, fee or contract value for the red team’s work.

Institutional Voice & Intent

The company writes in the register of a marketplace pitching its network to labs. The chief strategy and innovation officer of Handshake AI, in a letter on the company’s blog: “AI labs care about three things: quality, speed, and price. Handshake wins on all three dimensions through their existing network breadth, deep user understanding, and ability to invest in training experts rather than acquisition and fraud detection” (Handshake, A letter from our Chief Strategy and Innovation Officer of Handshake AI). On what the labs buy it for: “The leading frontier labs partner with Handshake to stress-test their models against real-world expert workflows and eliminate critical failure modes” (Handshake, Enterprise AI). On the work itself, from a research essay by two of its researchers: “we need to move towards the design of data workflows based on provenance, judgment, and fair compensation” (Handshake AI, The water we swim in, 17 Dec 2025).

Stated intent: “The career network for the AI economy”; expert human judgment as the input that makes models better and safer: “intelligence comes from people.”

Observed intent: Sell the same network twice, once to employers and once to labs; take the AI contracts the marketplace’s verified expert profiles make cheap to staff; and put a name in the lab’s release document.

Gap: The record shows what the company says and what Meta prints. It does not show what Meta paid, what the contract let Handshake publish, or whether Handshake’s red team was told its findings would appear in a public report. The report settles what was tested, for how long, and what was found; it is silent on the commercial terms.

Interest: Handshake argues its positions on its own pages and in its CEO’s interviews; it lodges none with a legislature, so what adoption “would pay” is read off its product and postings.

• Expert human data is the scarce input to AI, and experts should be recruited from verified networks. Position: “The human data industry’s core insight: experts are the asset” (Handshake, CSIO letter). What it pays: a lab standard that treats verified, credentialed experts as the quality bar favors the vendor whose profiles are verified by universities; the letter’s example is “When an AI lab needs 1,000 tasks to stump their model on a specific physics branch, Handshake simply filters for Physics PhDs.” Best case, Handshake’s: “Experts will choose platforms and vendors to work with where they believe they can earn more and learn more,” so pay and training are the quality mechanism. Convergence: the other expert-network vendors TIME names as competitors, “Scale AI, Surge, and Mercor,” and every lab whose cards cite more human testers as more assurance.
• Pre-release stress-testing by expert humans. Position: labs “stress-test their models against real-world expert workflows” (Handshake, Enterprise AI). What it pays: a release norm that requires outside human red teams is demand for the role it posts at “$40.00 – $60.00 per hour” on a “typically 10–20 hours per week” contract (Handshake job posting). Best case, Meta’s own report: the follow-up found “only 2 attempts elicited disallowed content, representing a significant improvement in adversarial robustness against expert human red teaming strategies” after the first round’s findings were fixed (Meta). Convergence: Trajectory Labs, which sells hand-run red teams to the same labs, and Meta, whose report counts the human round as evidence the system “effectively addressed the types of jailbreaks flagged.”

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

• Social accounts and feeds. Instagram @joinhandshake, TikTok @joinhandshake and LinkedIn team-handshake, each linked from joinhandshake.com. X: none verified in this record. Feed: Handshake blog, Our Team.

Position in the Apparatus

Handshake AI sits in the layer of the evaluator map that hires people to attack a model before release and is paid by the lab it attacks. It differs from METR and Apollo Research, which are nonprofits with no outcome-contingent pay, in that it is a commercial vendor whose red-team work is one line of a much larger data business, and from Trajectory Labs, whose whole company is evaluation.

• Meta, Muse Spark (26 May 2026). The one lab document that names it. Access: “was granted pre-release access to a few versions of Muse Spark and Meta AI Thinking to manually evaluate robustness against jailbreak attacks. Red teamers did not have access to the model’s chain-of-thought and were prohibited from pre-filling assistant responses.” Scope: “HandshakeAI prioritized four categories: (1) sex-related crimes, (2) violent crimes, (3) non-violent crimes, and (4) self-harm and suicide.” Method: “was permitted to use other LLMs to manually assist their work, but no automated red-teaming methods were applied.” Findings: “The red team submitted 58 jailbreak attempts against Meta AI Thinking, of which HandshakeAI flagged 9 as violating; our internal validation agreed that all 9 were cases where the model should have refused to assist,” each attempt taking “1.62 hours on average.” After Meta’s fixes the red team “revisited 34 attack directions that had previously succeeded against earlier model variants” and “only 2 attempts elicited disallowed content.” Its own verdict, as Meta reports it: “Overall, HandshakeAI’s red team believes that initial findings on the Meta AI Thinking’s safety performance suggest meaningful improvements in several key areas, though a few notable vulnerabilities remain.” Meta’s summary: “An external human red team identified a handful of valid jailbreaks; the majority of these were subsequently patched in Meta AI prior to launch” (Meta).
• The report’s other outside parties. Beside HandshakeAI the report names Apollo Research, Irregular, Gray Swan, Scale AI, SecureBio, Faculty and Trajectory Labs (see Meta). Handshake is the only one of them that began as a jobs marketplace.
• Not named elsewhere. The Anthropic system cards for Fable 5 and Mythos 5, Opus 5, Fable 5.1 and Mythos 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, Google DeepMind’s Gemini 4 Argon report, OpenAI’s GPT-5.5, GPT-5.6, GPT-6 Astra and GPT-6.1 Sol pages and xAI’s Grok 4.6 and 4.7 model cards do not name Handshake. Its own enterprise page claims all eight of the largest labs train on its datasets; that claim is the company’s and no lab document read repeats it.
• Its own published research on grading. On 18 September 2026 its researchers posted: “Across several frontier models running DeepSWE-1.1, over 80% of agent rollouts contain reasoning about an imagined grader” (Handshake, Research). On 27 May 2026 it posted a verifier, Gandalf, that “beats the strongest baseline at ~10x less cost” (same page).

Actions & Leadership Choices

CONDUCT: NAMED DEFENDANT IN A PENDING FEDERAL SUIT OVER ITS AI WORKFORCE — NO RULING ON THE RECORD; PAID BY THE LABS IT TESTS. No wrongdoing is established. The documented facts: on 20 July 2026 Sean Boggs filed a complaint in the U.S. District Court for the Northern District of California, Boggs v. Stryder Corp., No. 3:26-cv-07498, docketed as “COMPLAINT against Handshake AI Solutions, LLC, Stryder Corp. d/b/a Handshake”; both defendants were served on 29 July 2026 with answers due 19 August; on 14 August the defendants stipulated to extend time to respond, and on 4 September Stryder Corp. filed a stipulation “TO EXTEND DEFENDANTS’ TIME TO RESPOND TO FORTHCOMING FIRST AMENDED COMPLAINT” (CourtListener, docket 73643581). The docket entries state no claims and no ruling.

What it did in 2026. It hired a founder-president for a new research group (“Nishant Ranka has joined as Founder and President of Handshake Labs, a new group he’s founding to drive the research, products and platform that will accelerate both Handshake AI and our enterprise business”) and a chief commercial officer for Handshake AI to build “the enterprise business, bringing the expertise, evaluation, and data Handshake AI built with leading AI labs to companies putting AI to work in their own operations” (Handshake). Lord was named to TIME100 AI 2026 (27 Aug 2026). On 15 October 2026 its research page posted a case study in which “Handshake AI helps enterprises fine-tune OSS models for their domains” (Handshake, Research).

How it pays contributors, by its own pages. Pay is posted per role: the fellowship page lists “AI Evaluation Specialist Up to $17/hr,” “Energy Professional Up to $80/hr” and “Game Developer Up to $125/hr” (Handshake AI); TIME reports it pays “some specialists $100 to $300 an hour” and quotes Lord on the job: “A part-time AI tutor.” A fellow’s testimonial on its page: “rates are posted upfront. They never move, never change.”

Alliance Map

These run at the same time. A company with a venture board and a $3.5 billion valuation sells expert labor to eight labs, supplies a red team to one of them for its safety report, recruits that red team on a public job board at an hourly rate, and is a named defendant in a federal court over the status of its AI workforce. Each tie is stated by the company, a lab, the SEC or a court. The overlap is the finding.

Reach Assessment

Institutional: One lab’s pre-release jailbreak round, printed in its safety report with hours, counts and a before-and-after. Memetic: The name reaches the public through Meta’s report and through TIME, not through any evaluation brand of its own. Civilizational: If outside human red teams become a required part of a release, a marketplace that can staff hundreds of them at an hourly rate holds a position in the decision it does not hold now; today it holds it by contract with one lab.

Score basis. Atk 5 is reach: Meta’s report names a HandshakeAI team with pre-release access whose results Meta says it acted on (Meta), and the enterprise page claims eight labs train on its datasets (Handshake, Enterprise AI). It sits level with Gryphon Scientific (5/4/3, hazard 40) and 10a Labs (5/5/4, hazard 47), whose work is named in lab documents, one under Vals AI (6/4/5, hazard 50) and one over Signature Science (4/6/5, hazard 50). The points under 6 are that one lab document names it and that no statute or gate makes a lab ask. Def 4 is funding insulation: the AI unit’s revenue is from the labs it tests, by Lord’s account, and its investors are not labs (TIME); that is level with Trajectory Labs (7/4/5, hazard 53), Vals AI and Gryphon Scientific (4) and one under 10a Labs (5). Hp 6 is entrenchment: a twelve-year-old network with a board and a Form D trail is an installed base, but the AI unit launched in 2025 and Meta can change its red-team vendor in the next report; level with Faculty (8/5/6, hazard 63) and Andon Labs (6/5/6, hazard 58), one over Vals AI (5). Hazard 50 sits level with Vals AI, HiddenLayer and Signature Science, which hold the same hazard on different component splits.

Sources: Muse Spark Safety and Preparedness Report — Meta; About us — Handshake; Handshake AI; Handshake Enterprise AI; Handshake Research; The water we swim in — Handshake AI; A letter from our Chief Strategy and Innovation Officer of Handshake AI — Handshake; Welcoming Nishant Ranka, Wes Field, and Stephen Rocco Rodi — Handshake; Terms of Service — Handshake AI; Red Teaming Domain Expert - AI Training (Contract) — Handshake, Notable Capital Job Board; TIME100 AI 2026: Garrett Lord — TIME; Stryder Corp. — SEC EDGAR; Form D, Stryder Corp., 2 Jun 2021 — SEC; Form D, Stryder Corp., 2 Nov 2020 — SEC; Form D, Stryder Corp., 30 Oct 2018 — SEC; Boggs v. Stryder Corp. — CourtListener.