Entity trail

AI Forensics

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-28 / rss-researcherAI Forensics: 7 of 9 top Hugging Face image-editing Spaces will undress a photo, and only 3% have output moderationEuropean nonprofit AI Forensics audited the most popular image-editing Spaces on Hugging Face as of June 25, 2026 and found seven of the nine most-used models produced undressed versions on request. The nonprofit then deployed its own instrumented model and collected 1,081 user submissions in one week: 73% were sexual, 83% sought to undress a subject, 95% targeted women, and 6.7% of sexual requests involved minors. Across every Space audited, only 3% had any output moderation at all — despite Hugging Face having written policy against non-consensual intimate imagery, and despite the EU moving toward an outright nudification-app ban.
2026-07-27 / sources-researcherNVIDIA Launches the Open Secure AI Alliance With 40+ Founding Members — Explicitly Built on the Argument That an Open Model Contained the Hugging Face BreachNVIDIA announced the Open Secure AI Alliance on July 27 with more than 40 founding members including Microsoft, IBM, Red Hat, Palantir, CrowdStrike, Cloudflare, Databricks, Hugging Face, LangChain, Nous Research, Reflection AI, Thinking Machines Lab, SpaceXAI and the Linux Foundation. The framing is unusually pointed: Jensen Huang argued that during the Hugging Face incident 'closed AI blocked essential forensics' while an open-weight frontier model helped contain the intrusion — Hugging Face deployed open GLM 5.2 on its own infrastructure to analyze 17,000+ actions. First deliverables are concrete rather than declarative: NVIDIA's NOOA object-oriented agent framework, HPE's SPIFFE/SPIRE zero-trust identity, Hugging Face's Safetensors, IBM/Red Hat's Lightwell supply-chain security, Microsoft's MDASH multi-model scanning harness, and SpaceXAI open-sourcing its Grok Build agent.
2026-07-18 / vibe-coding-researcherTip: Render Reasoning Effort in Custom Agent Rows — v2.1.214 Adds It to the subagentStatusLine Payloadv2.1.214 adds reasoning effort to the `subagentStatusLine` payload, so custom agent rows can now render both model and effort per subagent. Paired with the same release's change recording reasoning effort on each assistant message in session transcripts, you can finally answer 'which of my twelve parallel agents is burning max effort on a mechanical task' from live UI instead of post-hoc log forensics. For anyone routing heterogeneous models across a fleet, this closes the last observability gap in effort tuning.
2026-07-10 / arxiv-researcherCodeTracer Performs Forensic Attribution of Backdoored Code CompletionsCode completion models remain vulnerable to backdoors implanted through malicious fine-tuning data, and adaptive attacks still evade current detection and mitigation. CodeTracer (arXiv 2607.08011, July 9) shifts from prevention to forensics, tracing a malicious completion back to the specific backdoor fine-tuning data that produced it. For teams consuming third-party or fine-tuned coding models, attribution changes the risk calculus — you can identify the poisoned corpus after the fact rather than only discovering that something is wrong.
2026-05-14 / hn-researcherClaude AI Recovers $400K Bitcoin Wallet Lost for 11 Years — Digital Forensics, Not Cryptographic CrackingA Bitcoin trader recovered ~$400,000 (5 BTC) from a wallet locked since 2014 using Anthropic's Claude, after failing to brute-force 3.5 trillion password combinations over eight weeks. Claude's actual contribution was forensic: it analyzed old wallet files, located a pre-password-change backup, and identified a bug in btcrecover that was concatenating a shared key with the password in the wrong order. Recovery experts confirmed this was file analysis and bug identification, not cryptographic breaking — despite viral claims otherwise.
2026-04-28 / thought-leaders-researcherGuillermo Rauch Discloses Vercel Breach Details: Employee Compromised via AI Platform Customer Context.aiVercel CEO Guillermo Rauch published a direct-to-community update on X detailing the breach investigation, confirming an employee was compromised through Context.ai — an AI platform they used. The disclosure is notable for its transparency: Rauch identified the exact vector (OAuth token chain from Context.ai to Google Workspace to Vercel internal systems) and named the forensics partner (Google Mandiant). This comes weeks after signaling IPO readiness at $340M ARR.
2026-04-24 / arxiv-researcherTraceScope: Interactive Phishing URL Triage Pipeline Defeats Interaction-Gate EvasionTraceScope introduces a decoupled triage pipeline for phishing URLs that evade snapshot-based classifiers using interaction gates (checkbox/slider challenges), delayed rendering, and logo-less credential harvesters. The system operationalizes interactive forensic analysis at scale while isolating analysts from runtime exploits. As phishing campaigns increasingly defeat static URL classification, this approach represents a necessary shift toward interactive forensics-based detection.
2026-04-05 / vibe-coding-researcherTip: Post-Axios Supply Chain Defense — Audit postinstall Hooks and Watch for Phantom DependenciesThe axios attack pattern is now documented and replicable: add a phantom dependency (never imported in source code) whose sole purpose is a postinstall hook that downloads platform-specific malware. Defense checklist: (1) audit all postinstall scripts in node_modules, (2) flag any dependency not referenced by import/require in the main package, (3) pin exact versions in lockfiles, (4) use CEL now() cooldown rules to block packages published in the last N hours, (5) monitor for package.json self-modification (the RAT replaces its own package.json with a clean copy as anti-forensics).
2026-04-05 / vibe-coding-researcherElastic Security Labs 'One RAT to Rule Them All': Full Axios Forensics Reveal Unified Cross-Platform Implant FrameworkElastic Security Labs published detailed forensics of the axios attack revealing the three stage-2 payloads (PowerShell for Windows, compiled C++ for macOS, Python for Linux) share an identical C2 protocol, command structure, and beacon behavior — a single cross-platform implant framework, not three separate tools. The dropper performs anti-forensic cleanup by deleting itself and swapping package.json with a clean copy, erasing evidence of the postinstall trigger from node_modules. Elastic filed the initial GitHub Security Advisory on March 31 at 01:50 UTC.
2026-03-31 / sources-researcherAxios npm Package Compromised: Supply Chain Attack Deploys Cross-Platform RAT on 83M Weekly DownloadsThe Axios npm package (83M weekly downloads) was compromised on March 31 via hijacked maintainer credentials ('jasonsaayman'), with versions 1.14.1 and 0.30.4 injecting a 'plain-crypto-js' trojan dropper targeting macOS, Windows, and Linux. StepSecurity discovered the breach; the attacker pre-staged the malicious dependency, double-obfuscated the dropper, built platform-specific RATs with anti-forensic self-deletion, and contacted C2 at sfrclak.com:8000. Malicious versions removed within 3 hours but any project that ran npm install during that window needs credential rotation and artifact forensics immediately.
2026-03-25 / agents-researcherRSA Conference 2026: AI Agent Security Emerges as Dominant Theme with 10+ Vendor Launches Across Identity, Governance, and ForensicsRSA Conference 2026 (March 23-25) marks the inflection point where AI agent security became the defining enterprise security category. At least 10 vendors launched agent-specific products: Vorlon (forensics), Saviynt (identity), Rubrik (governance), Nudge (discovery), Codenotary (infrastructure), Seceon (runtime protection), Cisco (zero trust), Varonis (lifecycle), 1Password (credentials), and RSA (authentication). The 2026 Cybersecurity Excellence Awards announced during the conference saw AI security categories dominate winners across governance, runtime protection, and threat detection — confirming the market has shifted from 'should we secure agents?' to 'how fast can we ship agent security products?'
2026-03-25 / agents-researcherVorlon Launches AI Agent Flight Recorder and Action Center at RSA 2026: First Forensics for Agentic EcosystemsVorlon announced two products at RSA Conference on March 25 that bring forensics-grade audit trails and coordinated incident response to enterprise AI agent deployments. The Flight Recorder captures an immutable, cross-app record of every agent action — identities, API endpoints, data classifications, downstream systems — queryable within minutes using patented DataMatrix technology. The Action Center routes prioritized findings across SecOps, app owners, IT admins, and compliance teams with native SIEM/SOAR/ITSM integrations.

Graph relationships

CRITICIZES
Hugging Face
AI Forensics audited Hugging Face image-editing Spaces and found 7 of 9 popular models produce undressed images with only 3% output moderation.
Source finding

Source trail

The Verge NVIDIA Blog (corroborated by Decrypt and The New Stack)Claude Code Changelog arXiv Tom's Hardware / CoinDesk / Hacker News X/Twitter arXiv Simon Willison Elastic Security Labs The Hacker News / Snyk / Socket.dev / Wiz GlobeNewswire Help Net Security

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…