AIxCC systems find 54 of 63 synthetic vulnerabilities, patch 43

AI system

System: AIxCC finalist cyber reasoning systems (Team Atlanta, Trail of Bits, Theori and four others)
Model: Not assessed
Release: Not assessed
Application developer: Seven finalist teams (Team Atlanta et al.), using LLM credits from Anthropic, Google and OpenAI
Role: Not assessed
Configuration: Not assessed
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: Injected flaws in real open-source projects with scoring for proofs and patches. Two years of funded development and frontier-lab model credits.
Human oversight: Not assessed
Relation to AI: DARPA, the organizer, reports the figures. The Record corroborates from the event. Stenberg's independent look at the curl tasks suggests the injected flaws were easy. That bounds how much the synthetic scores show.
AI contribution: DARPA, the organizer, reports the figures. The Record corroborates from the event. Stenberg's independent look at the curl tasks suggests the injected flaws were easy. That bounds how much the synthetic scores show.

Outcome identifiers: Not assessed
Benefit: found 54 synthetic vulnerabilities, patched 43, and found 18 real bugs for disclosure
Category: Reduced Risk
Status: Not assessed
Beneficiary groups: Society. Organizations.
Beneficiaries: Not assessed
Reach: Not assessed
Distribution and access: Not assessed
Duration: Not assessed
Expected?: Yes
Costs and harms: Not assessed

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: found 54 synthetic vulnerabilities, patched 43, and found 18 real bugs for disclosure
Comparison: the same competition's semifinal round (2024)
Results: Reference: 37% at semifinals. Result: teams identified 86% of the competition's synthetic vulnerabilities.
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Kind of value: average cost per competition task for the systems' bug reports and patches, which DARPA calls a fraction of the cost of traditional methods
Amount: $152
Currency and prices: Not assessed
Period and unit: Not assessed
Realization: Not assessed
Recipients and payers: DARPA (cost of running the competitors' cyber reasoning systems per task)
Gross or net: Not assessed
Included costs: Not assessed
Comparison: Not assessed
Method: Not assessed
Return and payback: Not assessed

Participants: Developer: Seven finalist teams (Team Atlanta et al.), using LLM credits from Anthropic, Google and OpenAI. Operator: DARPA (with ARPA-H), AIxCC Final Competition.
Jurisdictions: Benefit: Las Vegas, United States. Operator: United States.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Conditions for replication: Real-world codebases with harder, layered flaws than the injected ones. Maintainer acceptance of machine-generated patches.
Replication attempts: Not assessed
Later findings: The only later item found is curl maintainer Daniel Stenberg's 22 October 2025 review judging the injected curl flaws straightforward; no adoption, reversal or litigation was found.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Assessed benefits: Society. Organizations.
Depth: Not assessed
Reasons: Society: Aggregate competition results. No specific attack or deployment outcome documented. Organizations: Open-source maintainers received disclosures. No fix is individually documented in the sources.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: Independently reported

Kind: Not assessed
Related entries: Not assessed
Field: Cybersecurity / automated vulnerability discovery and patching. Government.
Purpose: Autonomous AI systems that find and patch vulnerabilities in open-source software underlying critical infrastructure.
Dates: Date basis unassessed: 2025 (precision unassessed). Updated: 21 October 2026.
Stage: Not assessed
Version: SHA-256: 01edde86def1…