The security industry spent decades making vulnerability discovery faster. We built better scanners, better research programs, better exploit intelligence, and better ways to describe what was wrong.
The backlog grew anyway.
Now the economics are changing again. Models can inspect code, test hypotheses, generate proofs of concept, and iterate with a speed and persistence that used to require a skilled research team. Agentic systems can also combine models with scanners, debuggers, shells, and other tools to operate across multiple stages of an attack.
Discovery is moving toward machine speed. Most remediation programs still move through a ticket queue.
That mismatch is becoming the defining problem in vulnerability management.
Key takeaways
AI is pushing vulnerability discovery toward machine speed, while most remediation programs still move through a manual ticket queue. That mismatch is the defining problem in vulnerability management.
The backlog predates AI. Researchers counted 48,185 published CVEs in 2025—roughly 132 per day—before misconfigurations, hardening gaps, and non-CVE issues.
Attackers reason across chains, not single findings. They need a workable path, not a critical vulnerability in isolation.
Security leaders should prioritize with context, separate routine volume from genuine exceptions, make validation part of remediation, and build an answer for problems without patches.
The goal: exceptions go to people, volume goes to the system.
The capability shift is already measurable
This is no longer a debate about whether language models might someday become useful for vulnerability research. In 2024, Google Project Zero reported that its Big Sleep agent found a previously unknown exploitable memory-safety issue in SQLite that existing testing infrastructure had missed.
The results have advanced quickly since then. In April 2026, Anthropic reported that its limited-preview Mythos model could autonomously find and exploit vulnerabilities in major operating systems and browsers. Its published examples included a 17-year-old FreeBSD remote-code-execution vulnerability and exploit chains that combined multiple weaknesses to bypass defensive layers.
Those are vendor-reported results from controlled research, not evidence that every attacker can autonomously compromise every target. The systems still make mistakes. Scaffolding, tool access, environmental context, and human direction still matter.
But the operational direction is clear: work that once depended on scarce human expertise is becoming easier to parallelize, repeat, and scale.
The volume problem existed before AI
AI did not create the vulnerability backlog. In 2025, researchers counted 48,185 published CVEs—roughly 132 per day. That excludes the misconfigurations, hardening gaps, exposed services, end-of-life systems, and proprietary applications that never receive a CVE.
No enterprise team can process that volume manually. More people may improve capacity at the margin, but staffing cannot make a ticket-based workflow scale at the rate software and findings are now being produced.
The same asymmetry is emerging in software development. Models can help teams write and review more code, find more bugs, and propose more fixes. That sounds like unqualified progress until the outputs reach the next human checkpoint. If an agent generates hundreds of pull requests and every change waits for manual review, the bottleneck has moved. It has not disappeared.
More findings do not reduce exposure. More patches do not reduce exposure. Risk closes only when the right change reaches the right system and the result is verified.
Attackers are starting to operate in loops
A model alone predicts tokens. A model connected to tools and wrapped in an agentic harness can do something more consequential: observe an environment, choose an action, evaluate the result, and repeat.
That distinction matters. Anthropic reported disrupting a 2025 espionage campaign in which attackers used Claude Code to support reconnaissance, vulnerability testing, credential harvesting, lateral movement, and data extraction. According to Anthropic’s investigation, AI performed an estimated 80–90% of the campaign’s tactical work while human operators retained control at several critical decision points.
A later analysis of 832 accounts banned for malicious cyber activity found model use across all 14 MITRE ATT&CK tactics. The important conclusion was not that every actor had become sophisticated. It was that scaffolding could help actors chain ordinary techniques into increasingly autonomous operations.
That changes what defenders should expect. An attacker does not need every weakness to be critical in isolation. They need a workable path: an entry point, a credential, a pivot, and access to something valuable. Models are getting better at searching for those paths.
The old operating model assumes time
Most vulnerability programs still follow a familiar sequence:
A scanner produces a finding.
Security enriches and prioritizes it.
A ticket goes to the system owner.
IT tests and deploys a change.
Someone closes the ticket.
Every step is reasonable. Together, they introduce delay. The team that identifies the risk often cannot touch the affected system. The team that owns the system has other operational responsibilities. Testing, approvals, change windows, and rollback planning are necessary because production changes can fail.
The answer is not to eliminate those controls. It is to encode them into an execution system that can act safely without requiring a new handoff for every routine change.
What security leaders should change now
1. Prioritize with context, not severity alone
Use CVSS as a description of intrinsic severity—not as a complete risk decision. Add known exploitation, exploit probability, exposure, reachability, asset criticality, available mitigations, and the consequence of a successful chain. CISA’s Known Exploited Vulnerabilities catalog should be a required input, not an optional feed. (For why severity scores alone fall short, see Attackers Don’t Care About Your CVSS Score.)
2. Separate routine volume from genuine exceptions
Do not put every remediation through the same manual path. Define scopes, confidence thresholds, approval requirements, maintenance windows, and rollback conditions. Let automation handle changes that satisfy those controls. Escalate ambiguous cases and high-blast-radius decisions to people.
3. Make validation part of remediation
A deployment is an attempt. Remediation is a verified outcome. Recheck the vulnerable condition after every change, preserve evidence, and roll back when the change fails or creates an unacceptable side effect.
4. Build an answer for problems without patches
Patch deployment covers only the cases for which a suitable vendor update exists. Misconfigurations, insecure defaults, end-of-life systems, custom builds, and emergency mitigations require a broader execution layer. A mature program must be able to reconfigure, uninstall, isolate, harden, or apply a compensating control—not just install packages.
Remediation has to become a continuous execution loop
The next generation of vulnerability programs will not be measured by how many findings they ingest or tickets they create. They will be measured by how reliably they turn a finding into a validated reduction in exposure.
That is the layer Furl is built to provide. Furl ingests findings from the tools already in place, deduplicates them, applies environment-specific remediation, validates the result, and preserves the evidence. Customer-defined scopes and approval gates determine where it can act. Failed changes can be rolled back.
You can see how that execution loop works on the Furl platform.
AI is not eliminating the need for human judgment. It is making it impossible to spend that judgment on every finding, every ticket, and every routine change.
Exceptions should go to people. Volume should go to the system.
If you want to see where continuous autonomous remediation fits in your current stack, book time with Furl.
Frequently asked questions
What is the exploit window in cybersecurity?
The exploit window is the time between when a vulnerability becomes known and when it is remediated on a given system. AI-assisted discovery and exploitation are compressing the attacker side of that window, while manual, ticket-based remediation keeps the defender side slow—widening the gap defenders have to close.
Why can't organizations just hire more people to fix the backlog?
With roughly 132 CVEs published per day plus the misconfigurations and hardening gaps that never receive a CVE, additional staffing improves capacity only at the margin. A ticket-based workflow cannot scale at the rate software and findings are now produced; the bottleneck simply moves to the next human checkpoint.
What is continuous autonomous remediation?
Continuous autonomous remediation is an execution model in which findings are ingested, deduplicated, remediated with environment-specific fixes, and validated automatically—inside customer-defined scopes and approval gates, with rollback when a change fails. Routine volume is handled by the system; ambiguous, high-blast-radius decisions are escalated to people.
About the author
Derek Abdine is the founder and CEO of Furl. He spent 25 years building the detection side of security, including as Head of Labs at Rapid7—where he led data science, threat intelligence, and security research, including the widely used Sonar internet-scanning project—and later as CTO at Censys. He holds five cybersecurity patents, has had research cited in the 2019 Economic Report to the President, and has briefed the White House and Capitol Hill. He built Furl to close the execution gap that detection alone never solved.