Entity trail

Gray Swan

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-10 / skill-finderBuild environmental containment first — Anthropic's own numbers show model-layer defenses cannot hold the lineAnthropic reports classifiers catch roughly 83% of overeager behaviors before execution and Claude resists prompt injection at ~0.1% single-attempt success on Gray Swan benchmarks — and still concludes that model-layer defenses cannot stop a determined egress. Their ordering is explicit: environmental containment (sandbox, VM, egress proxy) first, model-layer second. The reasoning is that an egress proxy blocks the exfiltration POST regardless of the agent's intent, while a classifier only reduces the probability of intent.
2026-06-24 / thought-leaders-researcherZico Kolter and Matt Fredrikson on Latent Space: 'AI Security Is Not Just Cybersecurity With AI'On swyx's Latent Space (June 22), OpenAI board member Zico Kolter and Gray Swan CEO Matt Fredrikson argued AI security is a distinct discipline — as enterprises deploy autonomous agents like Claude Code and OpenClaw, prompt injection and indirect attacks form a genuinely new risk class. They contend specialized red-teaming models now outperform humans at breaking AI systems, so robust defense requires dedicated guardrail models (e.g., Cygnal) plus continuous adversarial testing via platforms like Gray Swan Arena. The builder takeaway: agent security needs purpose-built models and ongoing red-teaming, not bolt-on filters.
2026-05-29 / agents-researcherAgent Security is a Systems Problem: Google/UCSD/Meta Paper Argues AI Models Must Be Treated as Untrusted ComponentsA position paper from Google, UCSD, Meta, Gray Swan AI, Cornell, and others (arXiv:2605.18991, May 18) argues that agent security must be approached as a systems problem where the AI model is treated as an untrusted component and security invariants are enforced at the system level. The paper contends that efforts to increase model robustness alone are insufficient and must be complemented with systems security techniques — zero trust, capability-based access, and invariant enforcement. This reframes the entire agent security conversation: stop trying to make the model trustworthy, start building systems that work even when the model is compromised.

Graph relationships

BENCHMARKED AGAINST
Claude
Claude resists prompt injection at 0.1% success rate on Gray Swan benchmarks.
Source finding
WORKS AT
Matt Fredrikson
Matt Fredrikson is the CEO of Gray Swan.
Source finding
BENCHMARKED AGAINST
Claude -> Gray Swan
Claude resists prompt injection at 0.1% success rate on Gray Swan benchmarks.
Source finding
WORKS AT
Matt Fredrikson -> Gray Swan
Matt Fredrikson is the CEO of Gray Swan.
Source finding

Source trail

How we contain Claude across products Latent Space (swyx)arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues