Security teams triage by severity score. Patch the criticals, work down the highs, and let the low-risk findings pile up, because each one alone doesn’t get an attacker much. Most organizations are sitting on an enormous backlog of exactly those — patches nobody is likely to apply soon, each one low-risk on its own. Vulnerability chaining is what turns that backlog into an attack path: string enough low-risk weaknesses together and the combination is a high-risk breach. Scoring systems acknowledge the gap: the CVSS user guide says CVSS is designed to rate individual vulnerabilities, and gives analysts separate guidance for scoring a chain exploited in one attack.
What changes with AI is speed. AI agents can already find previously unknown flaws: Google’s Big Sleep agent found an exploitable memory-safety bug in SQLite in 2024, which Google Project Zero called the first public example of its kind. Attackers already find individual vulnerabilities faster than they ever have, and tooling that can probe a system, spot disparate low-risk gaps, and assemble them into a working exploit path compresses a timeline that used to take a skilled human a long while. That undercuts severity-first triage, which was the thing making an impossible backlog tractable in the first place. The defensive answer is to get faster at detection and remediation across the whole backlog rather than better at ranking the top of it.
Follow one chain
A fictional demo service with two weaknesses:
- A lets an unauthenticated caller obtain a low-privilege session.
- B lets any logged-in session change another team’s records, because the write path never checks team membership.
Neither gets an outsider very far alone: A only grants a low-privilege login, and B needs a login to begin with. Together, A supplies B’s precondition, and an anonymous caller can change another team’s data. Fixing A closes this entry route but leaves B open to anyone already logged in; fixing B blocks the cross-team write but leaves A’s session for some other path. Breaking one link is real progress, but check the other weakness and alternate routes too.
To triage chains, map each candidate path: the starting access, each condition a step needs, the asset at the end, and the control that would interrupt it. Then check that control is actually enforced. Severity scores alone don’t capture this. CERT/CC’s SSVC deployer model instead weighs whether a flaw is being exploited, how exposed the system is, whether exploitation can be automated, and its impact.
Sources
- FIRST: CVSS v3.1 User Guide, section 3.4 Vulnerability Chaining — States that CVSS is designed to rate individual vulnerabilities and gives guidance for scoring a chain of vulnerabilities exploited in a single attack, listing each one alongside the chained score; checked October 26, 2026.
- Google Project Zero: From Naptime to Big Sleep — Reports an AI agent finding a previously unknown exploitable memory-safety bug in SQLite, which the team believes is the first public example of its kind; checked October 26, 2026.
- CERT/CC: SSVC deployer decision model — Prioritizes vulnerabilities by decision points including exploitation status, system exposure, whether exploitation is automatable, and impact, rather than by a severity score alone. Checked November 6, 2026.