October 16, 2026
AI Is Making Vulnerability Discovery Faster. Can Defence Keep Up?
Agentic AI and automated security testing are changing the rules of vulnerability management. The hard part is no longer finding…
By Confidence Ekeh
4 min read
Agentic AI and automated security testing are changing the rules of vulnerability management. The hard part is no longer finding weaknesses.
Imagine a security team that has spent weeks working through its vulnerability backlog.
Before the team can clear it, another round of automated testing identifies dozens of additional weaknesses.
Some are serious. Some may be duplicates. Others might look critical in a scanner report but pose less risk in the system's actual environment.
Now imagine that the tools identifying these weaknesses are becoming faster and more capable every few months.
The challenge is no longer just finding vulnerabilities. It is deciding what to fix first, and fixing it before the risk becomes an incident.
That challenge became more concrete on October 8, 2026, when Anthropic announced OSS Scanner, an opt-in service that gives eligible open-source projects periodic security scans by its most advanced AI models, at no cost.
Each report can include a self-contained reproduction of the issue, an explanation of the vulnerability and, where available, a candidate patch. But the reports are fully model-generated and delivered without human review. Maintainers receive findings faster, and they become responsible for deciding which ones are real and how serious they are.
The numbers behind the announcement are what caught my attention. Anthropic says its models have found more than 29,000 candidate vulnerabilities in critical open-source software over the past six months, but its team has only been able to manually review and triage about 6,000 of them.
Anthropic names the problem itself: it says it remains "bottlenecked on our human capacity to validate these findings." The organisation with some of the most capable vulnerability-finding models in the world is limited not by discovery, but by verification.
The early quality figures are strong. When expert penetration testers checked 97 critical and high-severity findings from an early version of the scanner, 85 (88%) met Anthropic's bar for disclosure. Of the remaining 12, 11 were real but duplicated other findings, and only one was invalid. Still, maintainers have told Anthropic that severity ratings can be inflated, or that the scanner sometimes misunderstands a project's threat model.
This is significant because open-source software supports a large part of the technology we depend on. Finding weaknesses earlier could help maintainers fix them before they affect everyone downstream.
But there is another side to this development.
Finding more vulnerabilities does not automatically mean better security
Traditional vulnerability management already involves several stages: discovering assets, identifying weaknesses, assessing risk, assigning responsibility, applying fixes and confirming that those fixes work.
Agentic AI can increasingly assist with several of these steps. Depending on the available tools and permissions, AI agents may help investigate findings, run security checks, trace potential attack paths, and suggest or test fixes. Anthropic says that while validating its scanner, some of the vulnerabilities it found could be chained into unauthenticated remote code execution exploits. That kind of attack-path work once required a skilled human tester.
Automated security testing also lets teams assess systems more often, rather than relying only on periodic assessments.
However, speed creates its own pressure. Anthropic notes that exploits can now be developed in minutes, and many maintainers are already choosing speed over certainty: the company says it has sent nearly 5,000 unverified reports to projects that asked for everything it had.
If discovery accelerates but validation and remediation stay slow, organisations end up with a growing backlog of findings. Teams could spend valuable time investigating duplicates or lower-priority issues while a genuinely exploitable weakness remains unresolved.
A severity score alone does not tell the whole story. Teams must also consider whether the affected asset is exposed, whether exploitation is practical, what business function it supports, and whether there is evidence of active exploitation.
What should security teams do differently?
I believe three practices will become increasingly important.
1. Prioritise risk, not the number of findings.
A critical-looking report should not automatically take precedence over every other issue. Teams need to weigh exploitability, exposure, asset importance and business impact. A report marked critical may not be critical in your environment.
2. Validate findings before acting on them.
AI-generated reports should be treated as evidence to investigate, not unquestionable conclusions. Security professionals need to reproduce the issue, understand the conditions it depends on and assess the proposed fix.
3. Make remediation part of the workflow.
Finding a vulnerability is only the beginning. Teams need clear ownership, realistic remediation deadlines, controlled testing of fixes and follow-up verification. Where AI agents assist with these steps, their permissions and actions must stay within clearly defined boundaries. An agent that can test a fix in production can also break production.
The real question is whether defence can keep pace
Agentic AI is changing the economics of security testing. Work that once required significant manual effort can increasingly be done faster and more often.
That creates an opportunity to catch weaknesses earlier. It also puts more pressure on the processes responsible for deciding what matters and making sure problems are actually resolved.
For me, the goal should not be to collect more vulnerability reports. It should be to reduce the number of exploitable weaknesses in the systems we protect.
As AI makes vulnerability discovery faster, what needs to change first in your organization: prioritization, validation or remediation?
Sources