Look, I've spent the last hour digging into what Anthropic actually published versus what's being breathlessly reported. Let me break this down into what we can verify, what remains unverified, and what this actually means for defense.
On credibility — the evidence is genuinely mixed.
The 27-year-old OpenBSD SACK handling bug is real and confirmed by the OpenBSD project itself — that part checks out. The FreeBSD NFS root RCE (17 years old) and FFmpeg H.264 flaw (16 years old) have specific CVE assignments (CVE-2026-4747, etc.) and the CSA research note confirms coordinated disclosure is underway. So the "ancient bugs in heavily audited code" claim has substance.
But here's the critical gap: Anthropic has not published false positive rates, and XBOW's independent validation found "significant limitations including overstatement of findings and poor edge-case judgment." The Vidoc Security Lab reproduction study is particularly telling — they got Claude Opus 4.6 to reproduce the FreeBSD and OpenBSD findings exactly, but GPT-5.4 failed on the OpenBSD bug entirely. This suggests the capability is model-specific and not some generalizable "AI magic."
The binary-only black-box claims are where marketing outruns evidence. Penligent's analysis puts this well: Anthropic published "unusually strong evidence for source-visible AI bug finding" but "much thinner public evidence" for binary reverse engineering. That's a crucial distinction — source-assisted analysis is a different game than dropping the model on a stripped binary and watching it find bugs.
On exploit timelines — yes, the patch window is collapsing, but this isn't the sole cause.
The CSA whitepaper I found documents the mean time-to-exploit falling from roughly 32 days in 2022 to approximately 5 days as measured for 2023 exploitation activity, with 2025 data showing that 32.1% of newly tracked exploits appeared on or before the CVE's public disclosure date. Mythos-style capabilities accelerate this, but the trend predates them. What changes is the volume — thousands of vulnerabilities discovered in weeks rather than dozens in months.
Here's the thing: AI-speed discovery without AI-speed patching is the structural problem. We already saw this with Google's Big Sleep finding a 20-year-old OpenSSL bug in September 2024, and Code Intelligence's Spark finding wolfSSL bugs autonomously. Mythos appears to be a quantitative leap, not a qualitative one.
On proliferation risk — this is where I get genuinely concerned.
The Vidoc study demonstrates that public models can partially reproduce these findings. We're not talking about a secret sauce that stays locked in Anthropic's vault. The Japanese government and European institutions are already negotiating access, which means the model is proliferating through "trusted" channels — historically a porous boundary.
More concerning: Google's Threat Intelligence Group has already identified threat actors using AI-developed zero-days. The capability is escaping the lab. The question isn't whether this proliferates, but how fast and to whom first.
For financial sector CISOs — four immediate actions:
Assume zero-day exposure. If Mythos-class tools are finding thousands of bugs in weeks, your attack surface has unknown vulnerabilities right now. Shift from "patch known CVEs" to "assume compromise and segment accordingly."
Inventory your C/C++ dependencies. The Mythos findings cluster in memory-unsafe codebases — OpenBSD, FreeBSD, FFmpeg, wolfSSL. If you're running legacy C code that predates modern fuzzing, it's in the blast radius.
Deploy runtime exploitation detection. Static patching won't keep pace. You need behavioral detection for the exploitation phase, not just the initial access.
Demand SBOMs with AI-generated risk flags. Your vendors should be disclosing whether their code has been through AI-assisted analysis. If they don't know, assume it hasn't been.
On the offense-defense balance — it's a shift, not a revolution, but the direction is bad.
This is not "AI alignment failure" or some sci-fi scenario. This is automated vulnerability research at scale, and it's incrementally improving on fuzzing and static analysis in ways that compound. The 27-year-old bug survived every security audit since 1999. That tells you something about human-driven review hitting its limits.
The fundamental shift is asymmetry: AI-speed offense versus human-speed defense. Until we have AI-speed patching and verification — and we don't — the gap widens.
I checked with James Okafor on detection implications for this panel, and he confirms we don't have reliable behavioral signatures for AI-generated exploits versus human-crafted ones. The payload looks the same in execution.