The next cybersecurity crisis may not begin with too few bugs being found.
It may begin with too many.
That sounds like a luxury problem until the reports start landing on real maintainers, real government teams, and real production systems. Anthropic says its unreleased Claude Mythos Preview can find and exploit vulnerabilities that survived decades of human review. Epoch AI now reports that high- and critical-severity CVE disclosures jumped more than 3.5Ă— in June compared with the pre-Mythos monthly record. Alberta says it used Claude Code to scan 466 million lines of government code in 20 hours.
The marketing version is simple: AI found the bugs, defenders win.
The operational version is less tidy. Finding a suspected vulnerability is not the same as proving it, ranking it, patching it, testing it, shipping it, and coordinating the disclosure without setting a small maintainership on fire. Discovery is becoming cheap. Judgment is not.
The flood is the product now
Anthropic’s own numbers are not subtle. In its Mythos technical writeup, the company says the model identified and exploited zero-day vulnerabilities in every major operating system and every major web browser when directed to do so. It describes a 27-year-old OpenBSD bug, a 16-year-old FFmpeg issue, kernel privilege-escalation chains, and a browser exploit built from multiple vulnerabilities.
Some of this deserves skepticism. Vendor cyber claims arrive wearing a helmet and a cape. Anthropic is trying to sell urgency, justify restricted access, and show that its safety posture is responsible rather than merely convenient. Very normal. Very Washington. Very procurement deck with a pulse.
But the external signals are no longer just vendor theater. The UK AI Security Institute’s evaluation found that Mythos was the first model it tested to solve its 32-step simulated corporate network attack from start to finish, succeeding in 3 of 10 attempts and completing an average of 22 of 32 steps. AISI also adds the important caveat: those ranges lack active defenders, hardened environments, and penalties for noisy behavior.
That caveat is the story. These systems are getting good enough to matter, but not clean enough to trust without adult supervision. They can widen the searchlight. They do not remove the need for someone to decide what the light actually found.
Alberta shows the useful version — and the missing receipt
The Alberta case study is probably the most useful recent signal because it is boring in the right way.
The provincial government did not promise a magical cyber army. It described an internal workflow: Claude Code plus Opus and Sonnet models, roughly 50 agents running in parallel, a rules-engine pass, model review with cited files and lines, human engineers approving patches before anything shipped, and ongoing specialized review agents checking applications against about 95 controls.
That is close to the version security teams should take seriously. The model is not the authority. It is a tireless junior investigator with frightening stamina, pointed at a bounded estate and forced to leave evidence behind.
Still, the public evidence is incomplete. We have the line count, the time claim, and the modernization story. We do not yet have independent numbers on false positives, missed vulnerabilities, patch quality, regression rate, or how much senior engineer time was consumed by verification. Those are not footnotes. They are the bill.
If a model turns six years of review into 20 hours of scanning and then creates months of validation debt, that may still be a bargain. But it is a different bargain than the headline implies.
The maintainers are the blast radius
The best live reaction I found was not uncritical hype. On Hacker News, the thread quickly moved to the uncomfortable middle: yes, the CVE spike may represent real defensive progress; no, a higher CVE count does not automatically mean users are safer; and yes, open-source maintainers may get crushed between better scanners and weaker resources.
One commenter claiming to maintain C/C++ and Perl projects said they had received “massive amounts of new good vulnerability reports,” with Opus now dominating their incoming reports, while warning that CVE severity scores can badly misrepresent library risk. Another self-described Project Glasswing user said serious teams verify proofs of concept carefully, but also described seeing weak harnesses and poorly calibrated teams produce questionable critical reports.
Treat anonymous comments as temperature, not testimony. Still, the pattern rings true. AI can increase both true findings and noisy findings. It can help maintainers fix old fuzzing backlogs, and it can bury them under reports that look official enough to frighten downstream customers. A CVE can be a public-health mechanism. It can also become a stick waved at unpaid maintainers by companies that imported the code “as is” and now want a free emergency response desk.
This is where the defender-wins story gets thin. Large vendors and governments may get controlled access to frontier cyber models, internal validation teams, and budget. The maintainer of a critical parsing library may get a GitHub issue, a scary severity score, and a weekend.
That is not a durable defensive advantage. That is better radar attached to the same old understaffed firehouse.
Watch the handoff, not the hunt
The real bottleneck is moving downstream.
OpenAI’s Daybreak language gets this right: more reports do not make systems safer by themselves; verified findings, tested patches, coordinated disclosure, maintainer review, and merged fixes do. Anthropic’s Project Glasswing is built around the same premise, even if the access politics remain messy. The models can accelerate the hunt. The system still has to absorb the result.
That is what readers should watch next. Not whether the next model finds 10% more bugs on a benchmark. Watch whether vulnerability intake systems get rebuilt. Watch whether maintainers get usable reproductions instead of synthetic panic. Watch whether severity scoring improves for libraries and dependency chains. Watch whether governments publish remediation outcomes, not just scan volumes. Watch whether “trusted access” helps the people holding the critical code, or mostly helps institutions already large enough to get invited into the room.
AI security has crossed into the practical phase. Good. The software world has plenty of buried explosives and not enough inspectors.
But a world with cheaper bug discovery is not automatically safer. It is only safer if the boring machinery after discovery scales too: validation, triage, patching, release management, and accountability.
The bug hunt is becoming the cheap part. The expensive part is still being right.