Cycode has launched Agentic Code Scanning, a system that decides which AI or rule-based engine reviews code and at what cost.
“Nobody got into application security to become a model economist,” said Lior Levy, Co-Founder and CEO of Cycode. “Point a frontier model at everything and the bill compounds with every commit, fastest on the code least likely to hide anything.”
Levy said pointing rules at everything leaves teams structurally blind to an entire class of vulnerability. Neither approach amounts to a full strategy, he argued, adding that the answer is a system that decides which engine runs where instead of simply a better model.
The launch responds to a set of questions security teams face once they adopt AI-assisted code review: whether to run a frontier model or an affordable one, whether to scan every commit or only nightly builds, whether to cover every repository or just the crown-jewel applications, and what to tell an auditor about which model produced which finding.
Cycode describes these as symptoms of one problem, namely relying on a single scanning method to do every job. The company positions the new capability as the fourth dimension of what it calls holistic code defense, sitting alongside three existing scanning methods it already runs in production.
Four code scanning dimensions reconciled into one view
Cycode runs four dimensions of analysis against the same code and reconciles the output into a single view of risk.
- Dimension 1 is deterministic SAST: pattern matching and taint rules built into the IDE, the MCP, pre-commit hooks, and a blocking pull request gate. It runs in under a second on a diff and returns identical results every time, which the company says makes it the only dimension fast and reproducible enough to block a merge.
- Dimension 2 is AI SAST. An LLM generates, updates, and customises the deterministic SAST configurations, so rule-based scanning improves across successive runs while its output stays reproducible.
- Dimension 3 pairs SAST with AI exploitability assessment. A rule engine finds candidate issues. An LLM then triages, ranks, and argues exploitability before a finding reaches a developer’s queue.
- Dimension 4, Agentic Code Scanning, sits above the other three and determines what runs, in what order, and on what budget. It plans from a threat model, ranks what it reads per vulnerability class, then attempts to disprove every finding it produces. Cheaper dimensions resolve known vulnerability classes first, directing model spend only where rules cannot structurally reach, so a vulnerability class does not get skipped simply because a budget ran out.
Findings get scored one at a time, Cycode says, but they are not exploited one at a time.
Agentic Code Scanning links related findings across files into a single multi-step attack path, carrying the data flow from an untrusted source to a dangerous sink. Cycode’s Context Intelligence Graph grounds each path in call relationships, ownership, and reachability instead of model speculation. Because a chain holds only while every link holds, remediation work targets the cheapest link to break instead of every finding along the path.
Those exploitability-qualified findings feed Cycode’s Agentic Workflows as a stream of events. Teams define a workflow once – setting the triggering events, the agent’s action sequence, confidence thresholds, and controls – then the workflow runs the moment a qualifying event occurs.
Findings are routed to the engineer who wrote the code and fixes are returned as pull requests validated against the original finding. Every run produces an audit trail recording what triggered it, what each agent did, and which boundary applied.
Benchmark results against six published CVEs
Cycode tested the approach against ten repositories spanning six languages, each pinned to an exact vulnerable commit. Six were real applications sitting at a published CVE. The other four were deliberately vulnerable applications carrying answer keys, which Cycode says its engine was never tuned against.
Against the six published CVEs, Cycode says the other vendors it tested returned zero valid catches. Its own Dimension 1 deterministic engine caught three of the six, all injection or traversal bugs with a code shape a rule could match. Agentic Code Scanning caught all six, according to the company.
Two of the six involved authorisation flaws: a broken access control issue in Gitea and a middleware authentication bypass in Next.js. No rule engine tested caught either one, including Cycode’s own.
An authorisation flaw is the absence of a check rather than the presence of a dangerous pattern, the company says, and that is why broken access control remains the OWASP number one risk regardless of how many rules a vendor ships.
The volume problem behind unranked findings
The taint rule that surfaced MLflow’s arbitrary file read fired 573 times across that repository. Two landed in the CVE file. The rest were scattered among 173 to 701 total findings per repository, none of them ranked, validated, or traced back to a root cause. Agentic Code Scanning returned one validated, root-caused finding per CVE.
Cycode frames the figures as results from single runs per repository on a reasoning-tier model, tested against a pinned public corpus. The company says it is publishing that corpus and its configurations so the results can be independently reproduced, a step that separates this from a closed vendor claim tested only in-house. Whether the same catch rate holds across a broader set of repositories, languages, or CVE classes remains untested in the material Cycode has released so far.
For security teams, the practical question raised by the benchmark is narrower than model selection alone. Two authorisation CVEs went uncaught by every rule engine tested, Cycode’s included, which points to a category of bug that pattern matching cannot reach regardless of how it is tuned.
The MLflow figures show a rule catching a real vulnerability inside 573 firings, with only two of those hits inside the actual CVE file. That distinction between detection volume and validated root cause is what teams will need to test against their own repositories once Cycode’s published corpus and configurations become available.
See also: Visa updates open-source VVAH tool with vulnerability remediation
Want to learn more about cybersecurity from industry leaders? Check out Cyber Security & Cloud Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the AI & Big Data Expo. Click here for more information.
Developer is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
