I’ve been trying to get real work done, but the Mythos disaster keeps landing in my inbox, so here’s a quick nod.
Remember Cybernews framing of the CVE-2026-5873 demo? The researcher “intervened when the model became stuck, redirecting it and providing debugging feedback.” That’s Clever Hans. The horse isn’t counting. The handler is cueing the horse. Clever Mythos.
Now Alisa Esage has posted claims of a Google Chrome exploit she found that proves the case. “Zero dupes so far” with Mythos. If Anthropic’s disclosures to Google covered the frontier, we should expect overlap with what she is reporting. She reports none.
That throws Glasswing under the big vulnerability bus. The April 7 Mythos rollout positioned the model as discovering apex-target bugs at scale. CVE-2026-5873 in V8, multi-chain renderer-plus-OS sandbox escape, the “too dangerous to release” label. If the model were operating where Esage operates, the two sets would intersect. They don’t.
Her business depends on the market for human-found Zero-days. Anthropic’s Mythos narrative attempts to price her out by running a cartel. She made a falsifiable claim anyway, timestamped, in public. Anthropic holds the disclosure set that would settle the dispute in one post. Let’s see it.
I write about this because I’ve spent a long career finding, reporting, triaging and managing vulns, and because the industry needs to take a hard look at the sudden threat to common sense from an AI corporation, lacking security expertise, trying to corner the security market. My business of putting out fires isn’t threatened by a sloppy fire-starter like Mythos. If anything they are generating a boom for security professionals who have to clean up their mess. I appreciate what Esage is saying and why, because she’s proving the mess.
Anthropic is clearly attempting to fence the entire security industry. Controlling what counts as a finding. What gets disclosed. Which CVE numbers get briefed to press. How the resulting narrative gets shaped.
Google receives the submissions. Reporters write the receipt. Nobody in the corrupted loop of the Church of Anthropic has to demonstrate that disclosed bugs sit anywhere near the actual exploitation frontier. The resolution mechanism sits inside the claimant. Centuries of enlightenment spent building external verification, abandoned for a press release.
Every time a CISO brings their entire team to me and asks “what is this AI speedup of vulns about?” I point them to the above analysis. Read and weep. It’s disinformation and narrative control by a get-rich-quick corporation, top-heavy with PhDs, that doesn’t understand risk management.
The next Anthropic step is probably to trick private equity muscle to act like enforcers and shove “enterprise-ready AI” into places it can’t work and doesn’t belong. This is radical capitalist predatory market behavior, completely decoupled from measures of engineering quality. It’s how you end up with a Tesla on your road, the worst piece of shit in automotive history killing hundreds of people with defective AI. Oops. Hard failure with actual deaths as a result, while some talent-less 2016 AI windbag vacuums up dollars for a decade on broken promises.
Look at the “call 0x41414141“.
She landed the hardest primitive class on a target where the mitigation stack is designed to prevent exactly that. That’s damning evidence. Zero dupes from someone demonstrating that level of control is a substantive technical smackdown on Anthropic’s capability. Esage posted a view of the unassisted frontier for a reason. Mythos is NOT on it.
Sorry Clever Mythos, you’ve been served.
