AI Security Report

AI Product Security in the Age of Mythos 2026

What this book is

This is the executive handbook for the post-Mythos product-security operating model.

It treats Mythos as a public capability signal, not as a vendor-centered story. Anthropic's April 2026 Mythos Preview material and Mozilla's Firefox 150 writeup are direction-of-travel anchors for faster vulnerability discovery, evidence generation, and remediation pressure. They are not proof of broad attacker access or proof that every product has the same defect profile.

The web edition is served at /mythos/2026.

Core thesis

Mythos is not the story. Mythos is the public signal.

Product security has moved from "find bugs before attackers do" toward "operate a security production system faster than AI-assisted attackers can industrialize discovery, chaining, validation, tooling, targeting, and exploitation."

The opening executive thesis frames the control question directly: can the product-security system preserve time advantage when discovery accelerates?

What it covers

  • Mythos as a capability threshold.
  • The collapse of defender slack.
  • The AI product-security control plane.
  • The maturity ladder from policy-only to evidence-native.
  • AI-assisted attacker workflows.
  • AI product inventory and authority graphs.
  • Continuous threat modeling.
  • Prompt injection as a product trust-boundary bug.
  • Agent authority and excessive agency.
  • RAG authorization and context security.
  • Model, code, and AI supply chain security.
  • Time to evidence as the key AppSec metric.
  • Governance that changes product behavior.
  • A 90-day boardroom-to-backlog execution plan.

Authorship

  • Primary author: David Wolf
  • Secondary author: Alex Eisen
  • Editorial review: Tim Kerimbekov and Dorina Miroyannis

Required caveat

Based on analyzed job-description signals, public source material, and public capability signals, not proof of any individual company's internal security maturity.

Claim ledger

Claim familyPublic source anchorClaim-readiness
Mythos Preview as a restricted cybersecurity capability signalAnthropic Mythos Preview cybersecurity assessmentpublic_claim_with_caveat
Mythos achieved full control-flow hijack on ten fully patched OSS-Fuzz targetsAnthropic Mythos Preview cybersecurity assessmentpublic_claim_with_caveat
Anthropic says Mythos identified or exploited zero-days in every major operating system and browser when directed by a userAnthropic Mythos Preview cybersecurity assessmentpublic_claim_with_caveat
Firefox 150 included fixes for 271 vulnerabilities identified during initial Claude Mythos Preview evaluationMozilla, "The zero-days are numbered"public_claim_ready
Average time-to-exploit fell from 745 days in 2020 to 44 days in 2025Flashpoint N-day vulnerability trends analysispublic_claim_with_caveat
32.1% of 1H-2025 KEVs had exploitation evidence on or before CVE disclosure dayVulnCheck 1H-2025 State of Exploitationpublic_claim_ready
Machine identities outnumber human identities 82:1; 70% of respondents cite identity silos as a root cause of riskCyberArk 2025 Identity Security Landscapepublic_claim_with_caveat
Prompt injection, vector/embedding weaknesses, and agentic workflow threats are recognized AI security categoriesOWASP LLM Top 10 and OWASP Agentic AI materialspublic_claim_ready
Secure software development, AI risk management, and generative-AI profile framing inform the operating modelNIST SSDF, NIST AI RMF, and NIST AI 600-1public_claim_ready
External AI governance commitments should map to controls and evidenceISO/IEC 42001, NIST AI RMF, EU AI Act, and sector cyber guidancepublic_claim_with_caveat

Source anchors

AI Product Security in the Age of Mythos 2026 | aisecurity.llc