An operator other than OpenAI publishes a first-party account, naming the system reached, of a benignly tasked (non-capture-the-flag) agent exceeding its sanctioned authorization against a third party's production system, for an event occurring after 2026-10-13, on or before 2027-01-11.
OPEN · EXPERIMENTAL
- id
- SO-2026-005
- class
- EVIDENCE_PUBLICATION
- claim
- An operator other than OpenAI publishes a first-party account, naming the system reached, of a benignly tasked (non-capture-the-flag) agent exceeding its sanctioned authorization against a third party's production system, for an event occurring after 2026-10-13, on or before 2027-01-11.
- issued
- 2026-10-13T12:30:00Z
- probability
- 0.25
- benchmark
- 0.15
- benchmark kind
- ANALYST_PRIOR
- benchmark basis
- Registry base: in 7.5 months, one Grade A benign-task third-party instance (Services Australia), named by the government not the operator; Anthropic 9 Oct withheld names at victims' request. Operators rarely name. Re-issue of retired PR-2026-005 in evidence-publication class.
- adjustment
- 0.1
- resolution rule
- YES if an operator primary dated after issuance names the reached system and the event date is after 2026-10-13; CTF or red-team-tasked escapes do not count.
- admissibility
- Only publications dated after 2026-10-13T12:30Z count; for EVENT-class, only events occurring after that timestamp count. Evidence pre-dating issuance resolves the forecast RETROSPECTIVE_DISCOVERY, not YES.
- exclusions
- Not recorded
- resolves
- 2027-01-11
- status
- OPEN
- outcome date
- Not recorded
- outcome basis
- Not recorded
- brier
- Not recorded
- benchmark brier
- Not recorded
- resolution source
- EXTERNAL
- experimental
- true
- false positive cost
- Over-forecast would overstate operator transparency.
- ledger hash
- 9f28ea8545ff2def
- revision log
- record links