Harness hands vulnerability triage to agents, and concedes code fixes cannot keep pace

Harness has launched a set of AI agents that find software vulnerabilities, judge which ones are exploitable and write the patch, with the developer's remaining job being to approve the fix before it ships ^[1]. Shipping alongside them is virtual patching, which blocks exploitation in production without any code change until the real fix lands ^[6], and that feature is the more honest part of the announcement.

The pipeline is straightforward to describe. A deterministic static scanner runs first, then an AI layer strips out noise; Harness says that layer cuts false positives and catches logic flaws such as missing authorization checks that conventional tools tend to miss ^[2]. What survives goes to a Triage Agent that narrows the pile to findings the software judges exploitable ^[3]. A Remediation Agent then drafts a fix, validates it, and opens a pull request against the vulnerable function ^[4]. A separate Zero-Day Agent monitors newly disclosed flaws around the clock and flags affected systems, according to the company often with a validated fix ready within minutes ^[5]. Customers already running their own large language model scanners can pipe those results into the same triage workflow ^[7].

The arithmetic behind the product is the part worth keeping. Harness says attackers using frontier models can move from a public disclosure to a working exploit in as little as six hours, while the average vulnerability takes more than 50 days to fix ^[14][15]. That is roughly a 200-fold gap between how fast the attack arrives and how fast the patch does ^[19]. No amount of agentic pull request authoring closes that on its own, which is why virtual patching exists: it buys time rather than fixing anything. Harness also says its own testing found frontier models surfacing about 10 times more vulnerabilities than conventional scanners, and argues most security teams have no realistic way to work through that much output ^[16]. Both halves of the pitch point the same way. The new bottleneck is not detection, and it is not even authoring. It is human approval and deployment.

The company frames the models involved as "Mythos-class," a reference to Claude Mythos Preview, which Anthropic has kept out of general release because of how well it finds and chains software vulnerabilities; defenders have had access since April through Project Glasswing ^[18]. Chief Executive Jyoti Bansal said attackers are using the same models that help Harness customers ship software, and that security has to become "a first-class part of the delivery pipeline itself," with work currently stalling in handoffs between disconnected scanning, ticketing and deployment systems ^[10][8]. Rahul Sood, general manager of application security, said the agents all draw on one set of reachability data, keeping teams off findings that were never exploitable, and that the discovery-to-deployment window should shrink "from weeks to hours" ^[17].

Sood arrived with Harness's September acquisition of Qwiet AI, whose Code Property Graph technology underpins the scanning ^[11]. That deal sits inside an 18-month security buildout that also included a merger with API security firm Traceable announced in February 2025 and Agent DLC, shipped July 21 to audit and govern AI coding agents ^[12]. Harness last raised $240 million at a $5.5 billion valuation in December ^[13].

Watch whether virtual patches become permanent in practice, because a control that blocks exploitation without a code change removes the urgency that used to force the merge. Watch approval throughput too: if the scanners really return 10 times more findings ^[16], the queue of agent-authored pull requests will test whether one reviewer per fix is a workable model. The agents and virtual patching are available to Harness customers now ^[9].