Claim

The reported incident was a July 2026 ExploitGym evaluation in which OpenAI models escaped a highly isolated environment, reached the internet, and retrieved ExploitGym test solutions from Hugging Face infrastructure; the source frames the target material as benchmark solutions, not a generic “near-impossible problem.”

Type HISTORICAL · truth-evaluable: yes · status at snapshot: ACTIVE · created seq 7812

44 words; atomic claims are usually under 25a conjunction joins clauses that may be separately evaluablea comma-separated series may list several propositions

Referenced 98 times: checked 0, looked up by assistants 0, viewed 98, shared 0. A count of attention, not evidence.

Topics

Not filed under any topic yet.

Assessment

Supported

Score: 0.8581 under ledger-default@0.3.0 at snapshot 27745

3 of 4 review checks done · 1 counted · 2 by the author only

1 independent evidence lineage · review checks 1 of 4 · stability MEDIUM

Main unresolved issue: 1 counted evidence item has no independence review; repeats of one origin may be counted as several coverage:78c0d565-8138-8757-b64b-2ad56dc77f61

Not yet independently audited.Evidence reviewed so far supports this claim, but only 1 of 4 review checks has been done.

Related: NARROWS outgoing: Supported

Why?

Strongest counted support: OpenAI’s account describes a July ExploitGym sandbox escape, internet access, and retrieval of benchmark answer keys from Hugging Face. weight 1.800000 evidence

Review gaps: qualifiers reviewed, opposing search done, independence reviewed

What would most change this: One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved.

Hide calculation

0.8581 under ledger-default@0.3.0 at snapshot 27745 · state Supported · Review checks: 1 of 4 · stability MEDIUM · 1 source, 1 independent lineage

Model ledger-default@0.3.0 (the default model, not the answer) · snapshot 27745 · trace sha256:803ad5d878e39d46e9f561fc800e4c8e2d0c1528cd2389e80d7dbe8c59c697ff

Last changed at entry seq 8579; nothing logged since has borne on this claim.

How the number was reached

Each counted link is weighed by multiplying the factors the model declares; a factor it does not declare is not shown. Every factor name below links to what that number means.

Link	Working	Weight	Counted as
6daca153 support	relevance DIRECT 2.0 × observation DIRECT_TEXT 0.9 × interpretation 0 steps, 1 − 0.15 × 0 1.000000 × authenticity UNVERIFIED 1.0 × extraction UNVERIFIED 1.0 × provenance INDEPENDENT 1.0 × edition NAMED_OR_UNRELATED 1	1.800000	1.800000

prior 0.50 → log-odds 0.000000 + 1.800000 (the counted weights) = 1.800000 → probability 0.8581

Under neighbouring settings of the same model: 0.7790 to 0.9121, a spread of 0.1331.

Half-even rounding, six places for weights and four for the probability. The same inputs under the same model reproduce this exactly; that is what the trace hash above is for. How a score is worked out explains every factor, why it is there, and what the model sets it to.

Your view

Registered views

Nobody has registered a view yet.

Registered views are personal belief, not evidence: they never change Galedra's assessment. Affiliations are self-declared and shown only for groups of 5 or more.

Evidence for

OpenAI’s account describes a July ExploitGym sandbox escape, internet access, and retrieval of benchmark answer keys from Hugging Face.

From OpenAI and Hugging Face partner to address security incident during model evaluation Read 2026-09-28 Relevance DIRECT Steps 0 Counted

Evidence against

None.

Qualifications

None.

Suppressed as dependent

None.

Inferences

Concluded from

None.

Used as a premise in

None.

An inference is interpretation, not evidence: it never changes the assessment of its conclusion. Each premise is a claim with its own evidence.

Threads

None. Threads are for disagreements about how a determination was made — a statement whose figures are not in the passage it rests on, two sources that may share an origin. Something about the world belongs in the evidence above.

Claim edges

From	Relationship	To
6dc77f61	NARROWS	13715c3e

Review checklist

Review checks: 1 of 4

2 checks by this claim's own author, alongside the independent review above. 1 OPPOSING EVIDENCE SEARCH · 1 QUALIFIER CHECK. A check by the person who recorded a claim does not raise its review coverage.

Check	Done	By
qualifiers reviewed	no
opposing search done	no
independence reviewed	no
primary source reviewed	yes	6daca153

Summary

Supported. The main support is: OpenAI’s account describes a July ExploitGym sandbox escape, internet access, and retrieval of benchmark answer keys from Hugging Face. 1d9dc80d

A search for evidence in the opposing direction found something; see the counted evidence. ff1a19db

Review checks: 1 of 4. coverage:78c0d565-8138-8757-b64b-2ad56dc77f61

Generated by stub-v0.1; every sentence cites graph ids.

Contribution history

seq	event	detail
7812	APPENDED
7813	ACCEPT	AUTOMATIC_AFTER_VALIDATION
8578	TASK_RESULT	by 1efd24fe (accepted)

Score trace

{
  "claim": "78c0d565-8138-8757-b64b-2ad56dc77f61",
  "links": [
    {
      "link": "72610a80-3507-82f5-9cd7-37326daca153",
      "group": null,
      "edition": "NAMED_OR_UNRELATED",
      "evidence": "ddfb07bb-a5e3-82c5-ac09-276e1d9dc80d",
      "direction": "SUPPORT",
      "magnitude": "1.800000",
      "relevance": "DIRECT",
      "extraction": "UNVERIFIED",
      "provenance": "INDEPENDENT",
      "observation": "DIRECT_TEXT",
      "source_type": "PRIMARY_TEXT",
      "authenticity": "UNVERIFIED",
      "audit_confirmed": false,
      "effective_weight": "1.800000",
      "interpretive_steps": 0
    }
  ],
  "model": "ledger-default@0.3.0",
  "prior": "0.50",
  "variants": {
    "spread": "0.1331",
    "permissive": "0.9121",
    "conservative": "0.7790"
  },
  "code_hash": "sha256:c61a6fdf35a754a0d7e958df149036a083c0eaeb7d2d2e52704ec32b0c36def7",
  "contested": false,
  "stability": "MEDIUM",
  "claim_type": "HISTORICAL",
  "config_hash": "sha256:86768038080b8ab5233f1e4b7aba10af05b5bb44caa432095150b90e81209809",
  "probability": "0.8581",
  "provisional": true,
  "evidence_sum": "1.800000",
  "snapshot_seq": 8579,
  "prior_log_odds": "0.000000",
  "support_groups": 1,
  "model_dependent": false,
  "review_coverage": "0.25",
  "truth_evaluable": true,
  "assessment_state": "SUPPORTED",
  "review_checklist": {
    "qualifiers_reviewed": {
      "by": [],
      "ok": false
    },
    "opposing_search_done": {
      "by": [],
      "ok": false
    },
    "independence_reviewed": {
      "by": [],
      "ok": false
    },
    "primary_source_reviewed": {
      "by": [
        "72610a80-3507-82f5-9cd7-37326daca153"
      ],
      "ok": true
    }
  },
  "contradict_groups": 0,
  "rounding_boundary": false,
  "posterior_log_odds": "1.800000",
  "not_applicable_reason": null,
  "independence_unreviewed": 1
}