VeriCUA: Verification as Agentic Exploration for Computer-Using Agents
The paper reports 90.5 percent AndroidWorld verification accuracy for VeriCUA versus 78.4 percent for a vision-language judge, with VeriCUA deciding on 82 percent of that sample.
AI security research / sources
Primary reports, advisories, papers and original records behind the Claims. Each record supports specific evidence rather than a whole Case by default.
1209 Evidence · Page 40 of 41
30 matching records on this page
The paper reports 90.5 percent AndroidWorld verification accuracy for VeriCUA versus 78.4 percent for a vision-language judge, with VeriCUA deciding on 82 percent of that sample.
Original Veritas paper constructs executable counterexamples against task graders and audits selected subsets of Agents' Last Exam, tau2-bench, DeepSWE and other benchmarks. It also tests generated replacement verifiers, with model- and harness-dependent results; sample findings are not whole-leaderboard error rates.
CVE-2026-122135 relied on Git signature-verification settings during startup.
CVE-2026-122133 described project-local executable lookup before Safe Mode checks.
CVE-2026-122134 used local Git filter configuration before trust approval.
vLLM advisory: EPS disguised as PNG reaches Ghostscript via Pillow
vLLM CVE-2026-114053 code_revision override advisory with an offline import-time PoC, requiring a dynamic AutoProcessor and trust_remote_code. Affected 0.7.3 through 0.30.x, fixed 0.31.0.
Advisory reproducing fatal engine behavior on legitimate balanced black/white image content.
Advisory on unbounded tokenizer chat templates monopolizing a shared server-renderer worker.
vLLM maintainer advisory GHSA-6757-hggm-hp2w shared-memory handle forgery
Advisory on browser-origin requests reaching loopback model serving without a configured API key.
Advisory showing a valid long audio input exhausting shared GPU memory in Phi-4 multimodal inference.
Advisory showing a distinct OpenAI-compatible endpoint missed request-template validation.
Study of attacker-written safety warnings in customer reviews and their effects on AI product recommendations.
CVE-2026-76066 analysis distinguishes hiding tools from users and enforcing permission at ToolNode execution.
VULONE analyzes exposed Cairn operator records, access outcomes and human-directed agent workflows; its project counts are not an independent victim census.
WADE: An LSTM-Based Anomaly Detection Engine for WebAssembly Applications
Washington Post October 26: Cathedral tests Anthropic external business systems
Forthcoming Section 15 describes consent, recognition of AI-agent-initiated transactions, customer monitoring and revocation responsibilities, with explicit applicable-law exceptions. The version is dated effective November 24, 2026, not already in effect on November 9.
Anthropic says it does not disclose customer or end-user information in response to government requests except through valid legal process, such as a subpoena or warrant, unless it believes an emergency may result in imminent physical harm or death and immediate disclosure could avert it. Anthropic also says it evaluates requests for legal validity, may reject requests that are broad or vague, and generally gives users notice unless prohibited or a rare exception applies. For API or enterprise customer data, Anthropic says it will generally ask the requester to contact the customer first.
Liu and Tsai quantify missing environment provenance among the 198 most-downloaded in-scope GGUF repositories and demonstrate partial bit-exact reconstruction.
Commit-specific harness analysis with selection-stage demonstrations.
Who Verifies the Tools? How Warped Formatters Mislead Agents and Their Verifiers, William Jeffries
Original WikiTrace paper evaluates linking condensed agent memories back to complete task trajectories across eight benchmark suites. Its reported utility gains are not a controlled test of forged traces, malicious memory injection or enterprise secret-retention safety.
Firsthand October 31 advisory reporting a separate affected agent wallet, task and hook references; aggregate loss figures remain allegations.
Wiz CIRT reconstructs a real cloud intrusion through an exposed Kafka UI, compromised cloud and IdP identities, and CI/CD or GitOps effects. The same IP scanned over 600 other environments, not 600 known breaches. The precise automation share and final exported data volume remain uncertain.
Research telemetry reconstructing two distinct compromised Bedrock credentials and broader incidents without asserting a common exploit.
Wiz traces malicious LiteLLM release artifacts, observed quarantine times and follow-on AWS/Bedrock reconnaissance, separately from vendor removal timestamps.
No matches on this page.