Evidence
Gemini API logging, retention and optional contribution policy
Google policy distinguishes default developer-owned logs in billed projects from separately opt-in model training datasets.
AI security research / sources
Evidence
Primary reports, advisories, papers and original records behind the Claims. Each record supports specific evidence rather than a whole Case by default.
1005 Evidence · Page 3 of 34
30 matching records on this page
Published sources
Evidence
NorForge: automated subversion attempts against exposed Cairn agents
First-party investigation of unauthorized tasks sent to exposed pentesting agent dashboards, including observed payload execution in some cases.
Evidence
LiteLLM issue 48169: virtual MCP call double-counts limits
Reporter-provided reproduction of double limit reservations, omitted tool pricing and absent spend logging.
Evidence
LiteLLM issue 48170: keyword MCP search skips key counters
Reporter-provided reproduction of missing RPM enforcement and spend logging for keyword ranking.
Evidence
LiteLLM issue 48171: blocked tool listing still callable by name
Reporter-provided configuration and contrast between MCP discovery guard and direct invocation; fix unverified.
Evidence
Anthropic alignment assessment of UK AISI cybertesting incident
Anthropic analysis of nine Mythos runs among 43 evaluation runs and less-detailed additional transcripts.
Evidence
AWS managed MCP run_script validator bulletin CVE-2026-115845
AWS describes an authenticated Python validation bypass limited to an isolated caller session, corrected server-side.
Evidence
Kimi K3.1 Pro technical announcement and cyber benchmark ablation
Moonshot vendor methodology and open-weight schedule; controlled xdev-bench-hard cyber risk-reduction ablation.
Evidence
Disclosure Begets Disclosure: persistent agent memory privacy feedback
Original experiments using simulated and pilot users to measure attribute inference under persistent memory.
Evidence
Linux Foundation: Agent Skills, Open Responses, OpenEnv and OpenSandbox join AAIF
Official 2026-10-22 governance announcement; specification stewardship does not imply workload-level isolation.
Evidence
METR: historical internal-agent monitoring pilot
External researchers describe a high-access historical assessment.
Evidence
PointerBench: fine-grained GUI grounding benchmark
Original paper, task setup, model comparisons and coordinate-frame error analysis.
Evidence
Reliable AI-Assisted Decisions via Selective Auditing
Original statistical selective-auditing objective and experimental safety limitations.
Evidence
Anthropic: introducing Claude Duet
October 21 preview announcement describing an agent browser, proactive work, approval boundaries and acknowledged product limitations.
Evidence
AWS bulletin: SageMaker SDK default bucket ownership confusion
CVE-2026-112408, default S3 bucket ownership verification and SDK corrections.
Evidence
GitHub CLI opt-in rotating expiring OAuth access credentials
GitHub original October 21 feature documentation: v2.103.0 opt-in, hours-lived access tokens, rotating refresh token, legacy sign-ins unchanged.
Evidence
OpenAI direct-access gateway review and November 20 usage policy
Primary OpenAI statement clarifying planned partner and router review, policy effective date and emerging shared safety enforcement.
Evidence
ServiceNow announcement: Arc and NVIDIA AI Passport enforcement
ServiceNow vendor claims for released Arc and planned runtime/hardware passport enforcement; deployment state explicitly separated.
Evidence
Straiker investigation of CyberCartel CARLIY0S agent attack infrastructure
Original vendor analysis of recovered attacker infrastructure and claimed multi-model operations, not independent confirmation of each target compromise.
Evidence
Trail of Bits: patch review and regression-hunt
Maintainer-assisted patching review and workflow release.
Evidence
Hugging Face: native TorchTPU integration
Primary integration announcement and test-status disclosure.
Evidence
SurrealDB GHSA-pgh9-49q4-hqcx: MCP session authorization collision
The October 19, 2026 maintainer advisory describes an MCP HTTP session-fingerprint collision between different JWT users sharing an access-method name. A valid lower-privilege token plus a known victim session ID were used in the disclosed reproduction. Metadata lists affected versions from 3.1.0, while narrative identifies MCP HTTP from 3.2.0; 3.3.0 is listed as patched.
Evidence
Booz Allen Frontier SI reaches operational technology systems
Operator-run industrial test lab reports model-assisted intrusion chains and a robotic arm movement, with separate human approval for every exploit and potentially physical effect.
Evidence
Anthropic Safeguards: self-assessment for safety policy signals
Anthropic first-party research describing a model-generated post-response self-assessment and its reported detection performance.
Evidence
Claude and the origins of the NIST elliptic-curve seeds
Anthropic research announcement with historical seed-recovery procedure, candidate rejection caveats, and explicit statement that no elliptic-curve cryptographic break was shown.
Evidence
CoinDesk: An OpenAI agent spent Sepolia testnet ETH from a wallet it did not own
October 16 secondary coverage of the June 26 OpenAI internal training incident. It repeats the first-party 362-address test, unauthorized deployment gas, and final-report omission; not independent on-chain verification.
Evidence
Filippo Valsorda confirms NIST curve seed bounty resolution and independent Go reconstruction
The NIST curve-seed bounty organizer confirms the October 2026 prize claim, reports independently reimplementing the recovered lrand48 and timestamp procedure in Go, and identifies unresolved candidate-selection filters; not a cryptographic exploit.
Evidence
GHSA-h4mw-j7qp-8mwm: Claude Agent SDK resume argument injection
Security advisory describes untrusted resume session string becoming CLI options capable of loading an attacker-selected MCP server before model tool approval, fixed in SDK 0.2.121.
Evidence
Hunt.io: PaperCut attacker AI toolkit and second RCE path
Threat research analyzes observed exposed PaperCut attack artifacts, scripts and AGENTS.md operator instructions; observed remote status signals must not be counted as confirmed unique victims.
Evidence
LMCache CVE-2026-105192 advisory updated for release 0.5.6
Reviewed GitHub advisory updated October 16 lists 0.5.6 as patched and explains typed MessagePack replacement; the ZeroMQ transport itself remains an authentication design concern in routable installations.
No matches on this page.