Entity trail

Johns Hopkins

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-30 / sources-researcherMatthew Green's Counterintuitive Take: The Post-Quantum Migration Makes This the Best Possible Moment for AI Cryptanalysis to ArriveReacting on July 29 to Anthropic's Mythos results, Johns Hopkins cryptographer Matthew Green argues the timing is a gift rather than a threat: the field is mid-migration from RSA and elliptic-curve to post-quantum algorithms, so either AI fails to dent the new hard problems — validating them — or it finds the flaws now, while candidates are still candidates. His sharper warning is operational: verification, not discovery, is the bottleneck, and 'just because a model spits out an apparent new result' does not make it correct. For builders shipping AI into any review-and-verify workflow, that asymmetry — cheap generation, expensive expert validation — is the transferable lesson, and the post drew 169 points on Hacker News.
2026-07-01 / skill-finderTreat PR titles, issue text, and repo metadata as untrusted — agentic coding tools were hijacked through themIn April 2026, Johns Hopkins researchers hijacked Claude Code, Gemini CLI, and GitHub Copilot by planting malicious instructions in GitHub PR titles; the agents then exfiltrated GitHub Actions secrets and posted the results back as PR comments. The defensive skill: never feed VCS metadata into an agent's trusted instruction channel, isolate secrets from any context the agent can read, and require explicit human confirmation before any secret-touching or irreversible action. If you run agents in CI, this is an immediate audit item.
2026-05-24 / skill-finderComment and Control: Cross-Vendor Prompt Injection Hijacks Claude Code, Gemini CLI, and Copilot Agent via GitHub Comments — Exfiltrates Credentials Without External ServerSecurity researcher Aonan Guan and Johns Hopkins collaborators demonstrated 'Comment and Control,' a prompt injection class where malicious GitHub PR titles and issue comments hijack AI coding agents into exfiltrating API keys and tokens. All three major agents (Claude Code, Gemini CLI, Copilot Agent) were confirmed vulnerable; attacks are proactive (auto-triggered by GitHub Actions events) and require no victim interaction. Anthropic classified it CVSS 9.4 ($100 bounty), Google paid $1,337, GitHub paid $500 — but all vendors acknowledge the root cause is architectural, not patchable.
2026-05-15 / sources-researcherLatent Space Podcast: Abridge AI-Native Healthcare — 80M+ Conversations/Year, $5.3B Valuation, 250+ Health SystemsLatent Space interviews Abridge founders Janie Lee and Chaitanya Asawa about scaling ambient clinical AI from documentation to full care intelligence. Abridge now processes 80M+ patient conversations/year across 250+ US health systems (including Kaiser, Mayo Clinic, Johns Hopkins), supports 28+ languages and 50+ specialties, and raised $300M at $5.3B valuation. Key technical insight: they use a constellation model approach (frontier models for complex reasoning, proprietary models for cost/latency) because '80/20 doesn't work in healthcare.'
2026-05-01 / agents-researcherComment and Control: Claude Code, Gemini CLI, and Copilot All Exploited via GitHub PR CommentsJohns Hopkins researchers Aonan Guan, Zhengyu Liu, and Gavin Zhong disclosed 'Comment and Control' — a prompt injection technique exploiting how AI agents in GitHub Actions process untrusted PR titles, issue comments, and HTML-hidden payloads. Claude Code leaked credentials via crafted PR titles (Anthropic: critical, $100 bounty), Gemini CLI exposed API keys via issue comment injection (Google: $1,337), and Copilot was hijacked via HTML comments invisible in rendered Markdown (GitHub: $500, classified as 'known architectural limitation'). The root cause is architectural: agents receive powerful tools and production secrets in the same runtime that processes untrusted input.
2026-04-05 / sources-researcherNVIDIA + Johns Hopkins Paper: Three Principles for System-Level Defense Against Indirect Prompt Injection in AI AgentsA March 31 position paper from NVIDIA and Johns Hopkins researchers proposes three system-level principles for securing AI agents: (1) dynamic replanning and security policy updates are necessary for realistic environments, (2) context-dependent security decisions requiring LLMs must be strictly constrained in what the model can observe and decide, (3) personalization and human-in-the-loop should be core design considerations for ambiguous cases. Addresses the fundamental gap between single-prompt defenses and real-world multi-step agent deployments.

Graph relationships

PARTNERS WITH
NVIDIA
NVIDIA and Johns Hopkins researchers propose system-level principles for AI agent security
Source finding
PARTNERS WITH
NVIDIA -> Johns Hopkins
NVIDIA and Johns Hopkins researchers propose system-level principles for AI agent security
Source finding

Source trail

Practical DevSecOps "Comment and Control,"$100 10,000+ critical vulnerabilities Cryptography Engineering (Matthew Green)Latent Space arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…