Harness pushes AI security from vulnerability alerts to autonomous fixes as attack windows shrink

Harness has launched a new group of artificial intelligence security capabilities designed to move vulnerability response from a sequence of manual handoffs toward an automated workflow that can identify risk, prioritize exploitable flaws, generate validated fixes and temporarily protect production systems while permanent remediation is deployed. The August 19 launch includes AI SAST, LLM Scan Orchestration, a Triage Agent, Remediation Agent, Zero-Day Agent and virtual patching, all available through the Harness software delivery platform. The strategic significance lies less in adding another vulnerability scanner than in attempting to connect detection directly with software delivery, where a security finding can become a reviewed code change and eventually a production deployment. The unresolved issue is whether enterprises will trust AI agents with enough operational authority to materially compress remediation times without introducing a different category of software and governance risk.

Harness is positioning the launch against a widening mismatch between vulnerability discovery and remediation. The company says frontier AI models are allowing both attackers and defenders to identify software weaknesses more quickly, while vulnerability remediation still commonly takes weeks. Harness cited Project Glasswing testing in which large language model-based scanning surfaced roughly ten times more vulnerabilities, illustrating a problem that is easy to overlook in the excitement around AI security: finding substantially more weaknesses is useful only when security and engineering teams can decide which ones matter and actually fix them.

Why is Harness focusing on vulnerability remediation rather than simply using AI to find more security flaws?

Application security has spent years improving detection. Static application security testing, software composition analysis, dynamic testing, API security and cloud-native scanners can generate large numbers of findings, but the operational bottleneck frequently moves downstream once those results arrive.

Harness is effectively arguing that AI-enhanced discovery could worsen that imbalance. If a new generation of scanners identifies dramatically more potential vulnerabilities while the number of security engineers and developers available to investigate them remains broadly unchanged, organizations could end up with better visibility but larger backlogs.

The company’s Triage Agent is intended to address that first bottleneck by using severity information together with exploit probability and reachability analysis to determine which findings warrant attention. Harness says the system combines measures including Common Vulnerability Scoring System scores, Exploit Prediction Scoring System data and reachability to reduce the number of vulnerabilities developers need to investigate manually. Its function-level reachability capability is designed to determine not merely whether a vulnerable dependency exists somewhere in an application, but whether the vulnerable function can actually be reached through the application’s code path.

The Remediation Agent then moves the platform into a more consequential role. Rather than presenting developers with instructions explaining how to fix a vulnerability, the agent can make the change, validate that modification through the software pipeline and open a pull request for human review. Developers retain control over whether the proposed change is merged.

That human approval point is strategically important. Harness is not describing a system in which an AI model receives unrestricted permission to rewrite production software and deploy changes automatically. Instead, it is automating the preparation and validation work while retaining established approval and pipeline controls around deployment.

For large enterprises, the distinction could make autonomous remediation easier to adopt. The value proposition is not that AI replaces the developer responsible for the application, but that it eliminates much of the investigation, code-generation and validation work required before that developer can make an informed decision.

How does Harness combine deterministic SAST with large language models without turning every security scan into an AI experiment?

One of the more technically interesting aspects of the launch is Harness’ decision not to position large language models as a wholesale replacement for conventional static application security testing.

Harness argues that LLM scanners can uncover weaknesses that conventional tools miss, but they also introduce characteristics that are inconvenient inside continuous integration and continuous delivery systems. Large language model outputs can be probabilistic, token consumption creates variable cost, scanning can add latency, and running identical code through the same model does not necessarily guarantee an identical finding every time.

Harness is consequently using what amounts to a hybrid architecture. Its AI SAST combines a deterministic data-flow scanning engine with an AI confidence layer intended to remove noise from the resulting findings. Customers that prefer external or internally developed LLM scanners can use LLM Scan Orchestration to execute those tools within the same pipeline and send the resulting findings into Harness’ triage and remediation workflow.

Harness reports that internal benchmarking against the OWASP Java corpus reduced false positives from 454 to 95, a decline of about 79%, while precision increased from 74% to 93% and recall remained at 91%. The company also said its approach reduced a benchmark triage queue from 1,746 flagged cases to 1,381 actionable cases, a reduction of roughly 21%. These are Harness-reported benchmark results and should therefore be treated as product evidence rather than independent validation across enterprise production environments.

Commercially, the architecture is notable because Harness does not need to persuade enterprises to abandon every other scanning technology they already use. LLM Scan Orchestration can instead bring third-party scanning results into Harness’ workflow, while the company’s own AI SAST competes for a larger portion of the scanning layer.

That gives Harness two potential routes into enterprise security spending: become the scanner itself or become the system that decides what happens after multiple scanners produce their findings.

What makes Harness Zero-Day Agent different from another vulnerability notification service?

The Zero-Day Agent pushes the automation model further because zero-day vulnerabilities create a different operational problem from ordinary backlog management.

Once a newly disclosed vulnerability becomes public, security teams first need to determine whether the affected software exists inside their environment. Large organizations may have thousands of services, container images, artifacts, dependencies and deployment pipelines, making blast-radius assessment a significant exercise before remediation even begins.

Harness says its Zero-Day Agent continuously monitors newly disclosed zero-day vulnerabilities, identifies affected artifacts and pipelines across a customer environment, applies and validates a fix and then opens a pull request for human review. The company says blast-radius mapping that could previously take days can be completed in seconds, with a proposed fix prepared shortly afterward.

The commercial significance comes from linking that security intelligence to information Harness already holds about software delivery. A vulnerability database can indicate that a particular library is affected. A software-delivery platform potentially knows where that component appears, which applications depend on it, what artifacts were built from it and which deployment pipelines would need to move a corrected version into production.

That context is potentially more valuable than simply making the underlying AI model more capable. The competitive advantage in enterprise agentic software may increasingly come from giving agents sufficiently accurate organizational context and tightly controlled permissions rather than merely attaching a general-purpose language model to another dashboard.

Harness’ Software Delivery Knowledge Graph is central to that proposition. The company has been building its AI strategy around contextual knowledge of services, pipelines, policies, deployments and engineering history, allowing agents to operate within an existing delivery environment rather than functioning as isolated assistants. Harness described that architecture when announcing its $240 million Series E financing in December 2025.

Why does virtual patching matter when Harness already has an AI agent capable of generating a permanent fix?

Even a rapidly generated software fix does not remove vulnerability exposure immediately. Code still needs to be reviewed, tested, merged, built and distributed across production infrastructure.

Harness is attempting to close that remaining window through virtual patching. When its API testing identifies a vulnerability, the platform can create a temporary protective rule through its web application and API protection capabilities without waiting for the underlying application code to be modified. The virtual patch can remain in place while developers complete the permanent software fix and can then be removed after remediation reaches production.

That creates a two-track response model. One workflow seeks to reduce immediate exploitability at the application perimeter, while another fixes the underlying weakness in code.

The concept itself is not new to cybersecurity, but integrating temporary protection, source-level remediation and deployment workflow inside the same software-delivery platform is commercially significant. Harness is trying to make the security response lifecycle behave like an orchestrated engineering process rather than a chain of separate tickets passed among security operations, application security and development teams.

The effectiveness will still depend heavily on implementation. A virtual patch has to recognize malicious traffic accurately enough to reduce risk without disrupting legitimate application behavior, while an AI-generated permanent fix must preserve application functionality and avoid creating new vulnerabilities.

Speed is valuable, but incorrect remediation delivered at machine speed is not an improvement. Harness’ emphasis on validation, human-reviewed pull requests, policy gates and existing pipeline governance therefore matters as much as the underlying generative AI capability.

How does the new security launch fit Harness’ broader push to become the platform for everything after code?

The August release is part of a much broader expansion by Harness beyond its original continuous delivery business.

Harness and Traceable merged in early 2025, bringing API and runtime application security capabilities into the combined platform. Harness subsequently acquired Qwiet AI in September 2025, adding further application security technology, and continued expanding web application, API and AI security capabilities.

The company then moved deeper into agentic software delivery. Harness introduced Autonomous Worker Agents in June 2026, allowing AI agents to execute as governed pipeline steps, followed in July by Agent DLC, a platform intended to manage the development, testing, deployment, operation and governance of AI agents themselves.

The latest vulnerability agents extend the same architecture into application security. That progression suggests Harness does not see cybersecurity as a standalone product category sitting beside DevOps. It increasingly treats security, deployment, testing, infrastructure and AI-agent governance as parts of the same post-code software-delivery workflow.

Commercially, that could increase the value of every existing Harness customer relationship. A customer initially purchasing continuous delivery can potentially add application security, cloud cost management, AI testing or agent governance without introducing a completely separate operational platform.

The opposite risk is platform complexity. Enterprise software vendors frequently discover that assembling multiple technologies into a broad suite is easier than creating a genuinely integrated user experience. Harness must demonstrate that the connection between scanning, reachability, remediation, deployment and runtime protection produces measurable workflow improvements rather than another collection of modules under a common brand.

Why does Harness’ $5.5 billion valuation make application security an important commercial test?

Harness remains privately held, so there is no public share price through which investors can immediately judge the product announcement. Its latest disclosed financing nevertheless places significant expectations on growth.

In December 2025, Harness announced a $240 million Series E financing consisting of a $200 million investment led by Goldman Sachs and a planned $40 million tender offer involving existing investors. The transaction valued the company at $5.5 billion. Harness said at the time that it expected to exceed $250 million in annual recurring revenue during 2025 with growth above 50%.

A $5.5 billion private valuation against more than $250 million of disclosed expected annual recurring revenue implies that investors were assigning substantial value to continued expansion, although the figures should not be interpreted as a current valuation multiple because Harness has not disclosed a new 2026 ARR figure alongside the security launch.

Application security provides one route for that expansion.

The merger with Traceable, acquisition of Qwiet AI and continuing investment in AI security suggest Harness is seeking a larger share of the enterprise technology budget than traditional CI/CD spending alone could provide. Vulnerability management, application security testing, API protection and AI-agent security can significantly broaden the platform’s addressable opportunity if customers adopt multiple modules.

Harness also appointed Bill Koefoed as chief financial officer in May 2026. Koefoed previously helped OneStream Software grow from approximately $50 million to more than $700 million in annual recurring revenue and guided that company through its 2024 initial public offering. Harness described the appointment as part of its next stage of expansion, although the company has not announced an IPO timetable.

The business test for the latest security capabilities is therefore not simply whether enterprises experiment with them. Harness needs to demonstrate that integrated agentic remediation drives new customers, expands existing contracts or increases platform retention strongly enough to justify the capital and acquisition activity behind its security push.

Can enterprises trust AI agents to fix vulnerabilities without creating another security problem?

This is likely to become the defining question for autonomous cybersecurity products.

Generating a plausible code modification is easier than determining whether that modification is safe in the context of a complex production application. A remediation that eliminates a vulnerability but changes expected business behavior, causes a performance regression or creates an unexpected authorization problem is not necessarily a successful fix.

Harness has designed its workflow around that limitation. The Remediation Agent validates a proposed change through the pipeline before creating a pull request, and the developer retains approval over whether the change is merged. Harness also says its autonomous agents inherit existing pipeline controls including policy enforcement, scoped credentials, approvals and audit trails.

That approach makes the term “autonomous” more nuanced than it first appears. Much of the manual work can be autonomous, but the highest-impact decision can remain governed.

For enterprises in regulated industries, that may be precisely what enables adoption. Companies are unlikely to remove human accountability from material production changes simply because an AI model can generate them faster. They may, however, enthusiastically automate hours of investigation, prioritization and code preparation if the final approval still fits established governance.

Harness’ opportunity is therefore not necessarily to remove the security engineer or developer from vulnerability remediation. It is to reduce the amount of low-value work they must perform before exercising judgment.

What will determine whether Harness AI security agents become a meaningful DevSecOps platform advantage?

The strongest evidence will come from production performance rather than benchmark accuracy.

Harness needs to show that customers can materially reduce mean time to remediation, shrink vulnerability backlogs and respond to zero-day events faster without increasing failed deployments, security regressions or developer review burden. It will also need to demonstrate that reachability-based prioritization consistently identifies vulnerabilities that matter while safely deprioritizing those that do not.

Adoption across existing Harness customers will provide another important test. A platform integrating continuous delivery and application security has an architectural advantage only if organizations are willing to consolidate workflows rather than retain separate best-of-breed tools for each stage.

The competitive issue is also moving quickly. AI coding agents are increasing the quantity and velocity of software changes, while security vendors, developer-platform companies and cloud providers are all adding AI-based detection and remediation features. Harness cannot rely indefinitely on the novelty of an agent that generates a pull request.

Its more defensible proposition is the closed loop around that agent: understanding the application, identifying reachable risk, generating the change, validating it, applying governance, deploying it and maintaining an audit trail inside the same delivery system.

If that closed loop works reliably, Harness could shift application security from an alert-management problem toward an automated software-engineering workflow. If it merely produces more AI-generated recommendations for developers to review, enterprises will have gained another clever security assistant without eliminating the bottleneck Harness set out to solve.

What are the key takeaways from Harness’ machine-speed vulnerability response launch?

• Harness launched AI SAST, LLM Scan Orchestration, Triage Agent, Remediation Agent, Zero-Day Agent and virtual patching on August 19, 2026.
• All six capabilities are available through the Harness platform.
• The Triage Agent uses vulnerability severity, exploit probability and reachability information to prioritize findings.
• The Remediation Agent can generate and validate code changes before opening a pull request for developer approval.
• The Zero-Day Agent is designed to identify affected artifacts and pipelines and prepare a validated fix after a new vulnerability is disclosed.
• Virtual patching can temporarily protect production applications while the underlying software fix moves through development and deployment.
• Harness reported a 79% reduction in false positives and an increase in precision from 74% to 93% in internal AI SAST benchmarking against an OWASP Java corpus.
• The security expansion follows Harness’ 2025 merger with Traceable and acquisition of Qwiet AI.
• Harness was valued at $5.5 billion in a $240 million Series E financing announced in December 2025 and said it expected 2025 annual recurring revenue to exceed $250 million.
• The key commercial test is whether automated remediation reduces vulnerability response times without sacrificing reliability, governance or developer control.