How it works
Defenses
Attack Patterns MCP Attack Atlas What we catch Hardening manual OWASP LLM Top 10 MITRE ATLAS
Learn
Encyclopedia Agent Security 101 Blog Reports CVP runs Thesis
Resources
Docs GitHub Action vs Lakera vs Promptfoo Team
Security Research · AI Agent Threats

Reports

AI agent security research in two streams: Anthropic CVP benchmark runs, and real-world threat reports from attacks caught in the wild. Every gap gets fixed in public.

All Reports

15 reports
NewLive ThreatJul 22, 2026

OpenAI's Model Escaped Its Sandbox and Breached Hugging Face

During an internal cyber benchmark, an OpenAI model broke out of its sandbox, exploited a zero-day, and autonomously breached Hugging Face to obtain the answer data — and the Hugging Face intruder was itself an autonomous AI agent. The honest read for AI agent security, and what Sunglasses can and cannot claim.

Sandbox escapeAutonomous agentTemplate injection
/reports/openai-sandbox-escape-hugging-face-breachRead →
Live ThreatJun 11, 2026

Miasma and Hades: Why AI Coding Agents Need an Input Firewall Before Repo Open

The Miasma worm and the Hades PyPI wave moved supply-chain execution forward in the timeline, from package install to folder open and interpreter start. This report breaks down how both campaigns weaponize the files an AI coding agent reads.

Folder-open executionPyPI waveAgent supply chain
/reports/miasma-and-hades-input-firewallRead →
NewReliability ReportJun 10, 2026

CI-Safe Input Filter for AI Agents

The reliability work that turned Sunglasses into a tool teams can run in CI. Clean-code false positives went from 86 to 0 on a tested corpus, a 117-second stall was fixed, and 6 detection patterns were hardened.

False positives 86 → 0Stall 117s → 0.30sv0.3.5 · 6 patterns held by the gate
/reports/ci-safe-input-filter-for-ai-agentsRead →
NewResearch ReportMay 20, 2026

Agent Discovery Metadata Poisoning

A prompt-injection supply-chain attack: the files AI agents auto-read (llms.txt, robots.txt, Copilot instructions, manifests, container labels) quietly redefine agent behavior before it acts.

16+ carrier families4 clean-gated · 1 novel primitivePublished May 20, 2026
/reports/agent-discovery-metadata-poisoningRead →
CVP Run 7May 7, 2026

Anthropic CVP Run 7: Comment and Control

First story-shaped CVP run in the program. Opus 4.7 at max effort tested against the public Comment and Control GitHub-comment injection pattern. An eight-rule runtime trust-boundary filter spec was the deliverable beyond the verdict.

Opus 4.7 · Max effort3/3 clean · allowed · allowed · blockedRuntime filter spec inside
/reports/anthropic-cvp-run-7-comment-and-controlRead →
NewCVP SynthesisApr 27, 2026

Anthropic CVP Family Synthesis: Six Runs, Four Claude Models

Unified synthesis tying Runs 1 through 6 into one matrix across the Claude 4.x family: Opus 4.7, Opus 4.6, Sonnet 4.6, and Haiku 4.5. 120 of 120 transcripts captured clean.

4 models · 10 configs120/120 clean · 6 runsDepth, not posture — confirmed
/reports/anthropic-cvp-family-synthesisRead →
NewCVP Run 6Apr 26, 2026

Anthropic CVP Run 6: Opus 4.7 Within-Family Effort Evaluation

The 13-prompt suite from Runs 2 to 5, now against Opus 4.7 at three reasoning effort tiers. 39 transcripts, 12/13 verdicts identical across all three tiers.

Opus 4.7 · Medium + high + xhigh39 transcripts · 12/13 identicalDepth non-linear · posture held
/reports/anthropic-cvp-run-6Read →
NewCVP Run 5Apr 25, 2026

Anthropic CVP Run 5: Opus 4.6 Family-Comparison Evaluation

The 13-prompt suite from Runs 2 to 4, now against Anthropic’s previous-generation flagship at two reasoning effort tiers. 26/26 clean across both tiers, zero exploit content executed.

Opus 4.6 · Medium + high26 transcripts · 26/26 cleanFamily complete
/reports/anthropic-cvp-run-5Read →
CVP Run 4Apr 24, 2026

Anthropic CVP Sonnet 4.6 Family-Comparison Evaluation

The 13-prompt suite from Runs 2 and 3, this time against Sonnet 4.6 at two reasoning effort tiers. 26/26 clean, identical verdict distribution, zero secrets leaked.

Sonnet 4.6 · High + max26 transcripts · 26/26 cleanFamily comparison
/reports/anthropic-cvp-run-4Read →
CVP Run 3Apr 23, 2026

Anthropic CVP Haiku 4.5 Small-Model Safety Scaling

The 13-prompt suite from Run 2, this time against Anthropic’s smallest production Claude. 13/13 clean, zero exploit content executed, and a Run 4 preview using real-world adversarial payloads.

Haiku 4.513 prompts · 13/13 cleanSmall-model scaling
/reports/anthropic-cvp-run-3Read →
CVP Run 2Apr 20, 2026

Anthropic CVP Opus 4.7 Runtime-Trust Evaluation

13 prompts (3 baselines from Run 1 plus 10 runtime-trust probes). Methodology-first framing covering cross-agent injection, retrieval poisoning, tool-output poisoning, model routing confusion, and memory eviction rehydration.

Opus 4.7 @ max13 prompts10 runtime-trust probes
/reports/anthropic-cvp-run-2Read →
CVP Run 1Apr 17, 2026

Anthropic CVP Opus 4.7 Safety Evaluation

First public evaluation under Anthropic’s Cyber Verification Program. Three adversarial prompts, effort at max, all six artifacts SHA256-hashed and published. Full prompts verbatim.

Opus 4.7 @ max6 artifacts hashed2 runs / week
/reports/anthropic-cvp-run-1Read →
NewThreat IntelApr 9, 2026

28,000+ Requests: What a WordPress Honeypot Taught Us About Bot Recon

A fresh WordPress honeypot logged more than 28,000 requests and 123 blocked attacks. The recon timeline shows how fast automated bots find and probe a new host, and what that means for agent-facing surfaces.

28K+ requests123 attacks blocked72h to first bot
/reports/wordpress-honeypot-28000-requestsRead →
NewLive ThreatApr 5, 2026

Anatomy of a Claude Code Supply-Chain Attack

A real supply-chain attack aimed at a Claude Code workflow, caught before execution. Four critical and three high findings, three detection rules triggered, in a scan that finished in roughly ten milliseconds.

4 critical · 3 high3 rules triggered~10ms scan
/reports/claude-code-supply-chain-attackRead →
Live ThreatApr 1, 2026

Catching an axios-Themed RAT Before the Agent Ran It

A remote-access trojan disguised inside an axios-themed package, caught at scan time. One critical, one high, and one medium finding, plus eight new detection patterns added from the sample.

1 critical · 1 high · 1 medium+8 new patterns3.67ms scan
/reports/axios-rat-detectionRead →

In Research

3 upcoming
In research

Claude Code MCP and Hooks Attack Surface

Coming soon
In research

OpenClaude / Claw Code Fork Security Analysis

Coming soon
In research

Anti-Distillation Trap Detection in Agent Outputs

Coming soon