Harness Ships AI Agents to Scan, Triage, and Patch Vulnerabilities at Machine Speed

Harness, the San Francisco software-delivery platform company, launched a suite of AI security agents on August 19, 2026 — AI SAST, agentic triage and remediation, a dedicated Zero-Day Agent, and virtual patching — built to compress the gap between finding a vulnerability and shipping a fix from weeks to hours.

The pitch rests on an asymmetry defenders keep losing. In Harness’s engineering post detailing the launch, the company cites the Edgescan 2026 Vulnerability Statistics Report’s average of 55 days to remediate a vulnerability, against attacker timelines (tracked by the Zero Day Clock) that have reached first exploit as little as six hours after disclosure. Even after a fix is written, the 2025 DORA Report puts commit-to-production at more than a week. Defenders are closing a window measured in hours with a process measured in weeks.

Frontier models now press on both sides of that gap. LLM-based scanning surfaces far more findings than traditional static analysis — partners in Anthropic’s Project Glasswing, which gives security partners access to the Claude Mythos Preview model to hunt vulnerabilities in critical software, saw roughly 10 times more vulnerabilities surfaced in testing. But visibility without processing is just a bigger backlog: Comcast, a Glasswing participant, found 44% of its critical- and high-severity LLM findings were false positives, and only 36% of critical and 33% of high-severity findings were confirmed exploitable.

“We’re at a point where the same AI models helping our customers ship software faster are also what attackers are using to find and exploit vulnerabilities faster,” said Rahul Sood, Harness’s General Manager of Application Security, in the company’s announcement. “The only way to close that gap is to make security a first-class part of the delivery pipeline itself.”

What the Agents Actually Do

The scan layer deliberately splits the difference between deterministic and probabilistic analysis, a notable design choice given the industry’s current enthusiasm for pointing LLMs at codebases. Harness’s AI SAST pairs a deterministic dataflow engine with an AI confidence layer that classifies each finding as confirmed risk, potential risk, or contextually safe. The company’s own benchmarking against the OWASP Java corpus claims the layer cut false positives by 79% (454 down to 95) and lifted precision from 74% to 93% while preserving 91% recall. On insecure direct object reference (a broken-access-control class that dataflow analysis structurally cannot see, since there is no malformed input to pattern-match), Harness claims 71% recall at 99% precision across a 390-case corpus in Go, Java, and Python. Those are vendor benchmarks, self-reported, but the underlying argument is sound: a probabilistic scanner that returns different findings on the same code twice, with a token bill per scan, is a poor fit for continuous CI/CD gating. Harness instead positions LLM scanning for one-off hunting and offers LLM Scan Orchestration for teams that want to run their own LLM scanners natively in-pipeline.

Downstream, the Triage Agent prioritizes findings by combining CVSS, EPSS, and reachability analysis: Harness now traces call paths to the specific vulnerable function, so a dependency is flagged only if that exact function is reachable. The Remediation Agent then writes a fix, validates it in the pipeline to avoid breaking builds, and opens a human-reviewed pull request. Developers keep merge authority.

The Zero-Day Agent is the piece aimed directly at the six-hour problem. It monitors for newly disclosed zero-days, identifies every affected artifact and pipeline across a customer’s environment (blast-radius mapping the company says drops from days to seconds) and produces a validated fix ready for review, skipping the triage queue entirely. Virtual patching covers the residual window: when API testing discovers a vulnerability, Harness deploys a protective rule on its web application and API protection layer with no code changes, shielding production within minutes while the permanent fix is written, then removes the patch once the fix ships.

All six capabilities are available now within the Harness platform, with the remediation and zero-day agents riding on the pipeline’s existing governance: policy gates, approvals, and chain of custody on every artifact.

Harness’s Build-Up to Agentic AppSec

The launch extends a security arc that began when Harness merged with API-security firm Traceable in February 2025, a deal that put Traceable co-founder Sanjay Nagaraj in charge of the combined company’s application security business. In July 2026 the company shipped Agent DLC, a governance layer for the AI agent development lifecycle: primitive scanning, an AI bill of materials, agent discovery, and an AI firewall enforcing policy against prompt injection and tool misuse at runtime. Days later it expanded its Kong partnership to push AI discovery and protection into Kong’s AI Gateway, inventorying every model, MCP server, and tool routed through it.

Unite.AI has tracked the same acceleration on the offensive side, from Claude turning a cyber benchmark into three real intrusions to IBM’s finding that AI played a role in one in four malicious breaches, and on the defensive side, OpenAI’s Daybreak cyber-defense models landing on Amazon Bedrock. Harness’s bet is that the fix pipeline, not just detection, is where machine speed has to arrive next.

Whether agents writing their own fixes earn enterprise trust will depend on the validation harness around them, a lesson reinforced when Copilot Autofix opened a shell injection in Snowflake’s CI/CD pipeline. Harness’s answer is architectural: the agents propose, the pipeline validates, and a human merges.