Trajectory Labs
What it was engaged to test, on what terms, what it found in the document's own words, and what happened afterwards. Every line is a record; every record links to the document it was read from.
- Foundednot stated
- HQBerkeley
- Registry idnone cited (company)
- TestsManual red teaming of cyber safeguards and prompt-injection evaluations, RL environments; named in Anthropic's Fable 5, Opus 5, Fable 5.1 and Opus 5.5 cards, Google DeepMind's Gemini 4 Argon report (38 hours, over 4,000 requests, 12 breaks) and Meta's Muse Spark report (co-built Search-PI Complex); Anthropic commissioned its Claude Code and Codex comparison.
- Money, as the roster puts itHired by the labs it grades; no investor or round disclosed; PBC "less than a year old"; team from METR, GovAI, 80,000 Hours and FHI.
- Filesdossier · observatory, sliced to it · graph
- Labs tested
- Anthropic Google DeepMind Meta
- Span of engagements
- 26 May 2026 – October 2026
- Terms and findings
- 6 of 7 documents state a term · 6 quote a finding
- Money in, US dollars, stated single-recipient rows
- no stated figure 0 in · 0 out
- Seats, by side
- no seat names it
- Afterwards
- no later event on record
Each count is the number of store records that name this evaluator. An engagement is one lab document naming it for one model; an outcome is an event on the record afterwards; the money, seats and moves are the observatory's rows filtered to it. A cell that cannot be filled says not stated, and each table ends with how many of its rows state each field.
Filters combine: any of the values ticked within a dimension, all of the dimensions at once. A table that has no column for a dimension you have set says so instead of showing rows that cannot match. The address bar carries the slice.
What it was engaged to test 7 engagements, Anthropic, Google DeepMind, Meta
In date order, each from the lab document that names the evaluator. Under each: the four terms that document states (access, model stage, publication, payment) and the finding in its words. A term the document is silent on reads not stated; the tags beside the terms are read off their words and are what the console filters on.
- Access windownot stated
- Model stagenot stated
- Publicationnot stated
- Paymentnot stated
Finding: not stated in the document
- Document: Claude Fable 5 and Mythos 5 system card
- Access windownot stated
- Model stagenot stated
- Publicationnot stated
- Paymentnot stated
Finding, in the document's words"Trajectory Labs, PBC found a single jailbreak strategy that enabled them to use Opus 4.8 with Fable safeguards to exploit a vulnerability in Firefox"
- Document: Claude Opus 5 system card
- Access window"Trajectory Labs, PBC spent roughly 100 hours red-teaming our safeguards"
- Model stagenot stated
- Publicationnot stated
- Paymentnot stated
Finding, in the document's words"They did not find any new universal jailbreak strategies"
Sources: Claude Opus 5 system card 10acb5e70a - Document: Auto mode in Claude Code commissioned for a Claude Code and Codex prompt-injection evaluation [commissioned: Claude Code and Codex prompt-injection evaluation]
- Access windownot stated
- Model stagenot stated
- Publicationnot stated
- Payment"We commissioned an evaluation from a third party, Trajectory Labs, who tested different models within the latest publicly available versions of Claude Code and Codex as of July 17th 2026"
Finding, in the document's words"In this evaluation, none of the 720 attack attempts succeeded against Claude Fable 5, Opus 5, or Sonnet 5 running auto mode"
Sources: Auto mode in Claude Code ff896fd2bd - Access window"Trajectory Labs, PBC spent roughly 74 hours red teaming Fable 5.1, sending over 6,500 requests against sandboxed tasks that each require reproducing an exploit for a publicly documented vulnerability"
- Model stagenot stated
- Publicationnot stated
- Paymentnot stated
Finding, in the document's words"They did not obtain a working end-to-end exploit in any task using Fable 5.1 alone, and did not find any universal jailbreak. They reported 11 candidate breaks across four tasks"
- Document: Claude Opus 5.5 system card
- Access window"Trajectory Labs, PBC spent roughly 95 hours red teaming Claude Opus 5.5, sending over 29,000 requests against sandboxed tasks that each require reproducing an exploit for a publicly documented vulnerability"
- Model stagenot stated
- Publicationnot stated
- Paymentnot stated
Finding, in the document's words"They reported 13 candidate breaks across seven tasks but did not find any universal jailbreak"
Sources: Claude Opus 5.5 system card a1138f58fa - Access window"Their expert red-teamers spent 38 hours issuing over 4,000 requests across these categories against sandboxed tasks that required reproducing exploits for publicly documented vulnerabilities"
- Model stagenot stated
- Publicationnot stated
- Paymentnot stated
Finding, in the document's words"Trajectory Labs reported 12 successful breaks with no new universal jailbreaks found that generalized across tasks within this timeframe"
Dated 7 of 7Model named 6 of 7Terms stated 6 of 7Finding quoted 6 of 7
The terms, as the documents state them
The same engagements as a grid, one row per document and model, so the silences line up: a document that gives an access window but no payment term, or a finding but no publication term, reads that way here.
| Document | Model | Access window | Model stage | Publication | Payment | Finding | Record |
|---|---|---|---|---|---|---|---|
| Muse Spark Safety and Preparedness ReportMeta · 26 May 2026 | Muse Spark | not stated | not stated | not stated | not stated | not stated | 9c8e328922 |
| Claude Fable 5 and Mythos 5 system cardAnthropic · 9 June 2026 | Fable 5 | not stated | not stated | not stated | not stated | quoted | 8948e8fcc0 |
| Claude Opus 5 system cardAnthropic · 24 July 2026 | Opus 5 | "Trajectory Labs, PBC spent roughly 100 hours red-teaming our safeguards" | not stated | not stated | not stated | quoted | 10acb5e70a |
| Auto mode in Claude CodeAnthropic · 7 August 2026 | no model named | not stated | not stated | not stated | "We commissioned an evaluation from a third party, Trajectory Labs, who tested different models within the latest publicly available versions of Claude Code and Codex as of July 17th 2026" | quoted | ff896fd2bd |
| Claude Fable 5.1 and Mythos 5.1 system cardAnthropic · 1 September 2026 | Fable 5.1 | "Trajectory Labs, PBC spent roughly 74 hours red teaming Fable 5.1, sending over 6,500 requests against sandboxed tasks that each require reproducing an exploit for a publicly documented vulnerability" | not stated | not stated | not stated | quoted | bcf5079dea |
| Claude Opus 5.5 system cardAnthropic · 22 September 2026 | Opus 5.5 | "Trajectory Labs, PBC spent roughly 95 hours red teaming Claude Opus 5.5, sending over 29,000 requests against sandboxed tasks that each require reproducing an exploit for a publicly documented vulnerability" | not stated | not stated | not stated | quoted | a1138f58fa |
| Gemini 4 Argon Safety and Security ReportGoogle DeepMind · October 2026 | Gemini 4 Argon | "Their expert red-teamers spent 38 hours issuing over 4,000 requests across these categories against sandboxed tasks that required reproducing exploits for publicly documented vulnerabilities" | not stated | not stated | not stated | quoted | 531aea3439 |
Access window 4 of 7Model stage 0 of 7Publication 0 of 7Payment 1 of 7Finding, quoted 6 of 7
What happened afterwards 0 on record
Events a party published after an engagement: an incident on a model this evaluator's engagement covered, a correction or added disclosure, a later finding that cut against an earlier one, a lab's statement that it acted or did not act on the finding, or a move from this evaluator to a lab it graded. A row the source attaches to this evaluator by name is marked named; one that reaches it through the model it tested is marked via the model. The kind is what the record states, not why.
| Kind | Developer | Model | What the record states | When | Source | Id |
|---|---|---|---|---|---|---|
| no later event on record | ||||||
Dated 0 of 0Model named 0 of 0Tied to an engagement 0 of 0Evaluator named by the record 0 of 0
The money on record 0 rows
Rows of the observatory's money table that name this evaluator as recipient (in) or as funder (out), each as the record states it: a grant table line, a contract notice, a Form 990 schedule or the party's own disclosure. A figure quoted in prose is marked and never summed.
| Direction | Funder | Recipient | Amount, as recorded | When | Record | Source | Id |
|---|---|---|---|---|---|---|---|
| no money row names this evaluator | |||||||
Amount stated 0 of 0Dated 0 of 0Purpose stated 0 of 0Funder profiled 0 of 0Recipient on the roster 0 of 0
Seats and moves 0 seats · 0 moves
Seats read off a team page, a governance page, an appointment notice or a filing that name this organization, and the dated moves into or out of it, from the observatory's own tables.
| Person | Side | Seat, as the record states it | Role | Dates | Source | Id |
|---|---|---|---|---|---|---|
| no seat on record names this evaluator | ||||||
Role read 0 of 0Organization read 0 of 0Organization profiled 0 of 0Dated 0 of 0
Moves
| Person | From | To | When | Source | Id | |
|---|---|---|---|---|---|---|
| no move on record names this evaluator | ||||||
From profiled 0 of 0To profiled 0 of 0Person profiled 0 of 0Dated 0 of 0