AI Cybersecurity

Securing AI systems, and AI used to attack and defend.

AI systems in DARPA's AIxCC final found 54 of 63 planted vulnerabilities in 54 million lines of code, patched 43, and found 18 real zero-days. On defense, the most robust model in a November 2025 prompt injection test fell 63% of the time when attackers got 100 tries.

AIxCC final results, Aug 2025: Planted, in scored code 63; Planted, found 54; Planted, patched 43; Real zero-days found 18; Real zero-days patched 11 AIXCC FINAL RESULTS, AUG 2025 Planted, in scored code 63 Planted, found 54 Planted, patched 43 Real zero-days found 18 Real zero-days patched 11 VULNERABILITIES
AIxCC first prize
$4M
Flaws Google Big Sleep found
20
CVEs in MCP software, 2025
99
Registry incidents
5

Evidence

AI found real vulnerabilities in 2025, and prompt injection still beats the best defenses

Documented cases: AI Cybersecurity
DateCaseCountryWhat happenedSource
Jul 2025Big Sleep finds SQLite flawUnited StatesGoogle's agent found CVE-2025-6965 before it could be exploitedSource
Aug 8, 2025AIxCC finalUnited StatesTeam Atlanta won $4 million; average 45 minutes per patchSource
Sep 15, 2025TC260 AI Safety Governance Framework 2.0ChinaFive-level risk grading and lifecycle technical measuresSource
Nov 25, 2025Gray Swan prompt injection testUnited StatesClaude Opus 4.5 at 4.7%, 33.6% and 63% attack success at one, 10 and 100 triesSource

Registry

Five registry incidents bear on AI cybersecurity

Registry incidents where this discipline was tested in practice. Each links to its full record.

Registry incidents linked to this page
IncidentDateDeveloperRegulator action
EchoLeak Turns Microsoft 365 Copilot Into a Data LeakJan 2025MicrosoftNone reported.
Stolen Drift Tokens Drain Salesforce DataAug 2025Salesloft (Drift)Affected companies filed state breach notices, including a Maine filing covering up to 28,414 people. No regulator action against Salesloft was found.
Nx Malware Turns AI Coding Tools Against Their UsersAug 2025Nx; Anthropic, Google and AmazonNone reported.
Chinese State Group Runs Espionage Through Claude CodeSep 2025AnthropicOn November 26, 2025, the House Homeland Security Committee asked Anthropic CEO Dario Amodei to testify.
OpenAI Test Models Breach Hugging FaceJul 2026OpenAISenator Josh Hawley wrote to OpenAI on September 9, 2026. The White House Accord on Super Intelligence, signed September 29, 2026, commits signers to controls that keep models from hacking systems in unintended ways. No enforcement action was found.

Rules

Security guidance is voluntary in the US; China mandates labeling and graded risk

Rules and frameworks: AI cybersecurity
RuleJurisdictionWhat it requiresStatus
OWASP Top 10 for LLM Applications 2025InternationalRanks prompt injection first among 10 risksNov 2024
AI Data Security information sheetUS and partnersData provenance, signing and poisoning protectionMay 22, 2025
AI Safety Governance Framework 2.0ChinaFive-level risk gradingSep 15, 2025
GB 45438-2025 AI content labelingChinaExplicit and implicit labels on AI contentIn force Sep 1, 2025

Status colors: green in force, blue voluntary standard or framework, amber adopted and phasing in, gray draft or pending, red withdrawn.

Record

Page Record

Datasets
ai_trust_evidence.csv (4 rows); registry incidents (5 linked)
Key sources
Big Sleep finds SQLite flaw
AIxCC final
TC260 AI Safety Governance Framework 2.0
Gray Swan prompt injection test
Last verified
October 2026