Tanium’s decision to join Anthropic’s Project Glasswing, announced this week, is a sensible move. The company is now using Claude Mythos, a frontier AI model built for vulnerability discovery, to scan its own production codebase before attackers do.
But the interesting story isn’t “AI joins the fight against vulnerabilities.” It’s a harder question: what happens after AI finds them?
The discovery to remediation gap
Anthropic’s own data from Project Glasswing gives a clear answer. Scanning more than 1,000 open-source projects, Claude Mythos identified an estimated 6,202 high or critical severity vulnerabilities, out of over 23,000 flaws of all severities. Of those, 530 have been disclosed to the maintainers responsible for fixing them. Only 75 have been patched, and just 65 of those have a public advisory confirming the fix.
That’s not a failure of the AI – the model did exactly what it was built to do, at a scale and speed no human team could match. The constraint is the human capacity to verify, triage, and patch what’s been found.
Anthropic’s own researchers put it plainly: progress used to be limited by how quickly vulnerabilities could be found – it’s now limited by how quickly they can be fixed. On average, a critical bug takes two weeks to patch once confirmed, with maintainers reported to have asked Anthropic to slow down its disclosure rate simply to keep up.
The same acceleration is happening on the offense
This gap matters because attackers are moving at the same accelerated speed. Research now puts the average time between a vulnerability being disclosed and it being actively exploited at under a day, down from roughly 2.3 years in 2019. Claude Mythos achieved a 72% exploit success rate and demonstrated the ability to chain multiple vulnerabilities into a single working exploit path without human guidance.
In July 2026, autonomous AI agents ran a four-day campaign against Taiwanese government infrastructure, mapping 21 connected systems, compromising 85 accounts, and exfiltrating more than 2,500 personnel records, confirmed by Taiwan’s Ministry of Digital Affairs. Notably, the attack didn’t rely on a single exploited CVE (Common Vulnerabilities and Exposures).[JP1] The agents built their entire operation from exposed authentication metadata, weak credentials, and misconfigured single sign on systems, the kind of exposure that exists in most enterprise environments today.
PointWire View
“Tanium joining Project Glasswing is a really positive move. Tanium technology is trusted to manage and secure millions of endpoints, many in highly critical environments, so using AI to proactively find vulnerabilities in its own code before attackers do makes complete sense.
It shows Tanium isn’t just talking about the potential of AI in security, it’s actively using it to strengthen the security of its own platform and, ultimately, better protect its customers.”
Richard Thompson – Chief Security Officer, PointWire
What this means for security teams
AI has changed vulnerability discovery, for defenders like Tanium scanning their own code, and for attackers automating exploitation. What hasn’t changed at the same pace is the human and process capacity to close the gap once something is found.
For security teams, the practical takeaway isn’t to wait for better tooling to solve this on its own. It’s to build the operational capacity to discover now, whichever side it comes from. A few steps stand out from the available evidence:
- Shorten patch testing and deployment timelines, rather than assuming the usual review cycle still fits the pace of disclosure.
- Treat foundational controls, such as multi factor authentication, network hardening, and comprehensive logging, as more important than ever, since they reduce risk independently of any single patch landing on time.
- Audit public facing authentication surfaces and developer documentation for information that autonomous agents can use to map your environment without needing a known vulnerability at all.
AI is transforming how vulnerabilities are found on both sides, however it’s not yet transformed how quickly organisations can act on what’s found. That gap is where risk lives, and closing it is still fundamentally a people and process problem, not a model to be deployed and left alone.
If your team is reviewing exposure across operational technology, remote access, or supplier-connected systems, PointWire can help identify where to focus first. Get in touch to discuss practical steps for reducing cyber risk.
Figures on Anthropic’s disclosure and patch rates are current as of its most recent published update and are likely to have moved since; readers should check Anthropic’s Project Glasswing page for the latest count before citing specific numbers.