Sysdig credits Anthropic's Mythos preview with 181 working Firefox exploits

The ratio carries more weight than the count. Two working exploits against Firefox's JavaScript engine from the previous frontier model, 181 from the Mythos preview, is roughly a 90-fold change in yield between generations ^[1][2][15]. Sysdig's post attaches no CVE identifiers to any of them, no disclosure dates, and no statement of which are fixed ^[14]. A remediation queue consumes advisories; a count consumes nothing.

The FFmpeg detail is the one that would change assurance practice if it holds. Sysdig says a 16-year-old flaw survived five million automated fuzzing iterations and did not survive the model ^[4]. Fuzzing hours are how a lot of teams, and a lot of their suppliers, demonstrate that a parser has been looked at. If iteration counts stop being evidence of coverage, the sign-off documents that cite them stop being evidence of anything.

Provenance matters more than usual here, because all of it is single-sourced ^[18]. The Firefox figures, the OpenBSD bug, and the overnight remote code execution written for an engineer with no formal security training ^[5] come from one publisher's blog post. The instrument Sysdig cites for time-to-exploitation collapsing to under a day in 2026, the Zero Day Clock, was built by Sysdig's own CISO, Sergej Epp ^[11]. The 9-to-12-month estimate for these capabilities becoming widely distributed is attributed to "the security research community," with no researcher named and no method given ^[10]. The one named technical judgment belongs to Nicholas Carlini, an Anthropic research scientist, who said in February that Claude Opus 4.6 was already better at finding vulnerabilities than human researchers including himself ^[6]. Mythos launched two months later ^[7][17].

Sysdig also says the capability was never trained for and emerged as a downstream consequence of general improvements in reasoning and code generation ^[8]. That claim has the longest reach of anything in the post, because it makes exploit generation a side effect of coding performance rather than a feature a lab can choose not to ship. Anthropic restricted Mythos through Project Glasswing, and Sysdig's own reading is that frontier capability proliferated regardless ^[9].

The briefing is more useful to operators than the exploit tally. The Cloud Security Alliance, SANS, [un]prompted and the OWASP Gen AI Security Project published "The AI Vulnerability Storm," reviewed by more than 250 CISOs and security leaders, Jen Easterly, Rob Joyce and Bruce Schneier among them ^[12]. It names 13 risks in four categories, maps them to NIST CSF 2.0 and MITRE ATLAS, and sets 11 priority actions, six rated critical with start-this-week deadlines ^[13]. Six of eleven is 55 percent of the action list on a seven-day clock ^[16], which is a schedule, and it arrives with no advisory behind it.

That leaves one number defenders own, the days between an advisory landing and the fix reaching production, and it reads the same today as it did last month. If discovery yield really moved 90-fold, that interval is where the added exposure lands.