In its first 14 days inside a gated preview, Anthropic’s new frontier model, codename Mythos, “wrote 181 working Firefox exploits,” while the previous state of the art produced two, according to Sila Ozeren Hacioglu of Picus Security. The same brief account says Mythos “surfaced thousands of zero‑days across every major OS and browser,” including a 27‑year‑old bug in OpenBSD, and that “over 99% of what Mythos found is still unpatched in production today.”
Anthropic’s Mythos and the scale of machine-speed discovery
Anthropic released Mythos to twelve partners under a gated preview in April 2026. The description in the Picus piece makes two blunt points: the model dramatically outpaced prior systems in producing working exploits, and most of its findings remain unpatched. The author frames this as a concrete event, not a projection: “That happened,” the piece emphasizes.
The 73‑second breach timeline and the 24‑hour patch lag
The article lays out a minute-and-a‑bit timeline of an AI-driven attack to make the dynamics tangible. “At second zero, the AI script kicks off. By second five, a CVE is exploited. MFA bypassed by twenty. Web shell dropped at thirty. Credentials dumped at forty‑five. By second seventy‑three, the compromise is complete.” In contrast, the defensive sequence — an alert, human triage, SOAR playbooks, ticketing and patching — often stretches into hours or a day: “The patch goes out the next day, twenty‑four hours after the breach that took seventy‑three seconds to complete.”
From months to hours: CVE‑to‑exploit compression
Picus summarizes the recent trend in weaponization timelines: a decade ago the median time from CVE publication to a working exploit was measured in months; by 2024 it had fallen to about 56 days, and by 2025 to 23 days. Current pairings from CISA KEV, VulnCheck KEV and exploit databases show a median delta of roughly 10 hours. As the article puts it, “Reversing a published fix into a working exploit is no longer a specialist craft; it's now a prompt.”
Why the “spaghetti handoff” defeats faster tools
The piece argues the bottleneck isn’t individual security tools but the handoffs between them — the “spaghetti handoff.” EDRs, SIEMs and vulnerability scanners may be fast, yet time is lost in human steps: Slack messages, copy‑pasted hashes, emailed PDF reports, tickets waiting for approval. Accelerating a single node doesn’t accelerate the whole graph; the time between detection and validated, evidence‑based remediation is where AI‑driven attackers now win.
Three pillars and a single under‑resourced axis: Identify, Protect, Validate
Picus sets out three fundamentals: Identify, Protect, and Validate. The third pillar, Validate, is split into two complementary practices. Defensive validation — Breach and Attack Simulation (BAS) — answers whether prevention and detection controls catch what’s actually hitting an environment and where residual risk remains. Offensive validation — autonomous pentesting — shows whether an attacker can chain exposures into a real path to critical assets. The article’s prescription is explicit: run both continuously in a loop so each informs the other, and then move that loop from “continuous” to “autonomous” so it operates at machine speed.
Autonomous validation, Picus’s approach, and the summit
Picus outlines autonomous validation as agents that read alerts, scope tests, run simulations, push fixes and write reports while SOC teams catch up. The company offers a guide with "12 operational recommendations security teams need" and "five actions for week one." Picus is also convening an Autonomous Validation Summit on May 12 & 14 with Frost & Sullivan, featuring practitioners from Kraft Heinz and Glow Financial Services and Picus CTO Volkan Erturk — a forum the piece says will unpack architecture, agentic workflows and operational realities.
How boards, security teams, and IT operations will respond
- Boards: The article reports that six months ago AI‑driven cyber risk was often delegated, but now “boards are treating it as existential and governing it directly,” freeing budgets for evidence‑based plans rather than “more of the same.”
- Security teams and technologists: They are urged to adopt both BAS and autonomous pentesting in a continuous, then autonomous, loop to answer “what’s actually getting through my controls today, and how far.”
- IT operations: The narrative points to the operational friction — tickets, approvals and manual playbook triggers — that turns a 73‑second compromise into a 24‑hour patch cycle; reducing those handoffs is presented as a practical priority.
Picus’s case, as presented by Sila Ozeren Hacioglu, is stark and specific: offense already runs at machine speed, and defensive confidence without validation is “guesswork at machine speed.” The tangible takeaway is operational, not philosophical — if the loop that proves and fixes vulnerabilities remains human‑paced, attackers running on AI will continue to find the answers before defenders do.
