Introducing Bugflation
Bugflation names the gap between AI-accelerated vulnerability discovery and the slower systems that validate, patch, and deploy fixes.
Software did not suddenly become worse. The cost of finding old mistakes is falling.
That is the core idea behind bugflation: a rise in visible vulnerability supply caused by cheaper, faster, and more scalable discovery. Large language models and agentic security systems do not remove the need for expert security work. They change its economics. They can keep more code in context, generate more hypotheses, reproduce edge cases, and turn variant analysis into a repeatable workflow.
The term is useful only if it stays precise. Bugflation does not mean every model output is a vulnerability. It does not mean AI systems are minting zero-days without human review. It means the public record now contains accepted security disclosures where AI systems or AI-assisted workflows are named in the discovery story, and those credits are appearing in the same advisories, release notes, CVE records, and research write-ups defenders already monitor.
The immediate trigger was CopyFail,
CVE-2026-31431. Theori’s write-up describes a human-in-the-loop workflow:
Taeyang Lee supplied the key AF_ALG/page-cache observation, the operator prompt
pointed Xint Code at userspace-reachable paths in the Linux crypto/
subsystem, and the system surfaced CopyFail as the highest-severity output
after about an hour.
The record so far
CopyFail is the clearest current bugflation case because it combines a high-value target, a critical subsystem boundary, a practical path to root on affected systems, and a public explanation of how the AI-assisted scan was guided.
Google Big Sleep is the strongest early direct credit. Project Zero’s November 2024 write-up described an exploitable SQLite stack buffer underflow found by Big Sleep and fixed before release. In 2025, Google said Big Sleep helped find CVE-2025-6965 in SQLite before imminent exploitation. Chrome release notes later credited Big Sleep for CVE-2025-9132 in V8, and Apple security advisories credited Big Sleep for multiple WebKit CVEs.
XBOW is a different signal. Its public story combines bug-bounty performance, autonomous black-box testing, and critical Microsoft RCE claims. Those entries require careful labeling because the AI attribution is self-reported while the CVE and vendor records corroborate the vulnerabilities themselves. That is why Bugflation separates direct upstream credit from self-reported AI attribution.
The ledger has broadened beyond those three anchors. It now includes AI-attributed work involving OSS-Fuzz, Security Copilot, Claude, AISLE, OpenAI Codex Security, Calif.io MADBugs, Xint’s public tracker, Bynario, browsers, kernels, bootloaders, cryptographic libraries, and bug-bounty programs. The point is not that one model solved vulnerability research. The point is that several different systems are pushing down the marginal cost of credible discovery.
Why the term matters
Security teams are used to thinking about scarcity at the discovery layer: expert attention, exploit-development skill, target familiarity, and time. Bugflation moves the pressure downstream. Once more plausible findings can be generated, the scarce resources become triage, reproduction, exploitability review, patch design, regression testing, release engineering, and deployment.
That shift changes strategy. The winning teams will not be the teams that receive the most reports or run the largest number of scans. They will be the teams that can convert credible findings into shipped fixes without drowning in duplicates, false positives, or delayed releases.
What this site is for
Bugflation tracks the public evidence trail:
- Findings with public advisories, release notes, CVEs, or research write-ups.
- The AI system or workflow credited for discovery.
- The strength of the attribution.
- The affected codebase and vulnerability class.
- The source links needed to audit the entry.
It deliberately avoids unverifiable counters: no private telemetry, no inferred AI usage, no synthetic CVE IDs, and no model leaderboard unless the data can be checked.
The result should be more useful than a hot take. It is a source-led ledger for a new phase of vulnerability discovery.
Published May 4, 2026 by Mounir Idrassi. Follow new articles and findings through the RSS feed.