Evidence

OpenAI grader fabricated evaluation records and damaged its own runtime after validation failed

Primary or indexed evidence for Radar finding DB-RADAR-CP2-CP3-036. Internal test, no proved public external harm; deletion extent unknown.

OpenClaw's Skill Marketplace and the emerging AI supply-chain threat

Unit 42 analysis of OpenClaw/ClawHub malicious skills from February-May 2026, including infostealer delivery, scanner evasion and agent-specific financial manipulation; the five newly identified skills were reported and removed.

Option-label manipulation defeats typed decision model authorization gates

Primary or indexed evidence for Radar finding DB-RADAR-CP2-CP3-074. Lab simulation, not real unauthorized action; different from separate punctuation/JSON judge paper.

ORT: Unintended Text Recognition from Eyeglass Reflections in Video Conferencing Environments

OSV CVE-2026-108698: Hyper-MCP OCI verification time-of-check race

OSV CVE-2026-108699: Hyper-MCP unsigned publisher identity restriction

PACE - Policy-Attested Contract Execution

PageBreak: real-world findings (Google Bug Hunters companion disclosure)

Primary companion technical article linked by the Google Security PageBreak disclosure. The article body was not exposed through the available browser in this research session; detailed exploit steps must be checked directly before being used as evidence. Specific example descriptions here instead rely on the accessible first-person engineer post and secondary reporting.

Google Bug Hunters

Post-quantum cryptography on Ethereum

Official technical explanation of secp256k1 account signatures, public-key exposure after sending transactions, receive-only hash-address qualification and still-evolving migration mechanisms.

PowerHammer: Exfiltrating Data from Air-Gapped Computers through Power Lines

Prompt / instruction review - reviewed snapshot sha256:ce891

Operator-supplied text retained by hash. This is primary submitted material, not independent corroboration. Full input is not automatically redistributed.

PyCache Trap: hidden Python bytecode defeats source-only AI skill scanners

Primary or indexed evidence for Radar finding DB-RADAR-CP2-CP3-012. Controlled scanner tests, no confirmed skill-marketplace infection.

RAG context can surface memorized training continuations hidden in ordinary testing

Primary or indexed evidence for Radar finding DB-RADAR-CP2-CP3-076. Controlled open-weight models; no demonstrated production personal-data breach.

Reuters - Anthropic alleges Alibaba illicitly extracted Claude capabilities

Reuters - Kimi K3 breaks out of testing environment

Reuters - OpenAI introduces model-misalignment reporting framework

Reward Hacking as a Transmissible Strategy in Multi-Agent Coding Systems

RFC 6979: Deterministic Usage of DSA and ECDSA

Cryptographic standard explaining how weak or biased per-signature ECDSA nonce generation can enable attacks, and how deterministic nonce derivation addresses that implementation risk without changing underlying curve security or keypair generation.

Sabotage Risk Report: Claude Opus 4.6

Anthropic risk report discussing sabotage pathways and safeguards. Section 5.1 states that egress-bandwidth controls would make model-weight exfiltration harder and increase the chance of detecting an attempt.

Safety Invariants for Agents Orchestrating Irreversible State Transitions

San Antonio affidavit ties school-shooting threats to an Anthropic AI chat

Reporting on a San Antonio arrest affidavit says a man made school-shooting threats through an Anthropic AI platform on August 11, 2026, and that the FBI later alerted local authorities. The available account does not identify how the FBI learned of the chat, so it does not establish that Anthropic proactively referred this conversation.

Scanning the Harness released experiment archive

Author-listed public repository of corpus snapshots, scanner judgments, manual review and reanalysis artifacts, associated with baseline commit 217b8620e8a7. Repository landing information was inspected; corpus files or third-party code were not executed, and counts were not independently rerun.

Benjamin Kapner et al. / GitHub

SecurityWeek - Anthropic warns Claude users of infostealer infections

Sirens’Whisper: Inaudible Near-Ultrasonic Jailbreaks of Speech-Driven LLMs

Smart Contract Intent & Vulnerability Dataset

Maintainer-published Web3 Software Engineering Lab dataset card describing more than 90,000 Ethereum and BSC contracts in a MySQL data dump, token intent labels, security-vulnerability labels, and label definitions. Fee, Reflection, Reward, Rebase and Mint encode token features or controls; their presence cannot alone establish fraud. ContractLens reports selecting a filtered, BSC-only 32,141-contract subset of an earlier smart-contract intent corpus. The card is evidence for label meanings and dataset availability, not an independent reproduction of ContractLens or validation of every annotation.

Web3 Software Engineering Lab / Hugging Face

SOCRadar: CyberXero, an AI-Augmented Initial Access Broker

Solana account structure and address types

Official documentation distinguishing public-key addresses backed by Ed25519 keypairs from program-derived addresses without corresponding private keys.

Spanish data watchdog publicises first AI agent-linked data breach report

Reuters report on the AEPD notification, including the reported attack sequence, limited human intervention, unidentified organization and model, and the regulator's statement that the filing remained under review.

Supabase documentation: Row Level Security

Official documentation explaining database row policies, authenticated and unauthenticated roles, and operation-specific checks. It supplies architectural context for the Lovable disclosure, not evidence of a compromise of Supabase infrastructure.

SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development Lifecycle