Snyk launches Evo Agentic AppSec with security agents that find and fix vulnerabilities

What Snyk announced on 15 October

Snyk's newsroom release on Evo Agentic AppSec is dated 15 October 2026 and frames the product as security that moves at the speed agents write code. CTO Manoj Nair's quoted line is blunt: a model grading its own homework is a conflict of interest with a dashboard, not a security programme.

A companion product page for Evo Agentic AppSec and a first-look blog on Remediation Agent and Malicious Code Defense fill in the loop: continuously understand an application, find exploitable issues, ship validated mergeable fixes, and block malicious packages before they land.

Snyk says the agents run on the deterministic engines more than 4,800 customers already use. That claim is the company's; AI Lookout has not audited customer counts or engine coverage.

Why Snyk says AppSec has to change for agent-written code

The press release argues that critical and high-severity issues detected across Snyk customers grew 10x in the past year, and that product data showed roughly eight new vulnerabilities introduced for every one remediated. Separately, Snyk says its researchers responded to 18 high-profile malware-focused attacks in 2026 and project a more than 500% annual increase in malicious packages published to open-source registries. Those figures are Snyk's analysis of its own telemetry and research; they are not independent industry totals.

Snyk also cites environment and skill audits: one in 12 developers running MCP servers had a high or critical finding in nearly 10,000 environments checked, and 36% of nearly 4,000 audited agent skills had at least one security flaw. Again, those are vendor-run samples.

The company's VulnBench report, quoted in the release, says LLM security reviews report a finding that disappears on the next run 49.7% of the time. VulnBump, also quoted there, says LLMs alone produce a mergeable fix only 59.6% of the time. On the same mergeable-fix standard, Snyk says Evo Remediation Agent raised that rate to 90%. The product page separately mentions a 94% improvement in mergeable fix rate for SCA issues, a 14% improvement for SAST and 60% lower token cost versus a vanilla LLM. Treat every percentage as Snyk's own benchmark until someone else reproduces it.

The agents in Evo Agentic AppSec

Evo Code Scanning Agent pairs frontier-model reasoning with Snyk's static analysis to catch logic the company says pattern matching misses, especially broken access control and authentication flaws in AI-generated code.

Evo Dependency Firewall is meant to block malicious and out-of-policy packages at install time, including slopsquatting names that coding agents hallucinate. Evo Agentic Security Review delegates a full repository review: build an application and threat model, check whether controls match credible scenarios, and return evidence-backed findings with remediation guidance outside ordinary scanner coverage.

Evo Remediation Agent runs the fix loop across new code and backlog, starting with Snyk Open Source issues at launch. Fixes that might break the build are queued for approval rather than merged. Evo Secrets Detection covers leaked credentials across traditional and agentic development from IDE through CI. Each agent, Snyk says, draws on an intelligence layer with more than a decade of fix-outcome data and can call reachability, exploitability and breakability analysis as a shared resource.

How verification is supposed to stay independent

Snyk's core design claim is separation of generator and validator. Agents call Snyk Code, Snyk Open Source and other engines as tools; those engines also verify what the agents propose. Continuous Offensive Security, generally available since August according to the release, is cast as an independent judge that attacks the running application the way an adversary would.

Prevention inside a coding agent needs no human sign-off in Snyk's framing, because an agent already authorised to write code should fix a vulnerability it just introduced. Remediation of code that may already be running requires human approval before it ships. Developers move toward approving agent-proposed fixes; AppSec leads move toward setting bars, governing agents and auditing decisions.

That 'humans on the loop' posture is closer in spirit to programmes that keep models inside a harness than to unconstrained auto-merge bots. For related approaches to AI-driven vulnerability hunting with containment, see our coverage of Chainguard's Mythos Harness waitlist and Singapore's cybersecurity strategy focus on AI harnesses.

What is generally available versus preview

Availability is uneven by design. Evo Code Scanning Agent, Evo Dependency Firewall and Evo Agentic Security Review are experimental previews for select Snyk customers. Evo Remediation Agent is in open preview. Secrets Detection, Evo Agentic Development Security (ADS) and Evo Continuous Offensive Security are generally available. Malicious Code Defense is described in the first-look blog as private preview.

Snyk's docs changelog already lists Remediation Agent under early access and notes broader Evo surfaces such as a hosted MCP server, Continuous Offensive Security GA and Agent Behavior Governance. Organizations are told to talk to their Snyk representative or visit snyk.io for access. There is no public self-serve price list in the materials we reviewed.

Snyk Studio, which the release says more than 1,600 customers have adopted, remains the conversational remediation surface in the command line. ADS was previously released to prevent vulnerabilities inside the coding agent before code is emitted. Agentic AppSec is the programme that runs after code exists.

How this sits next to other AI security programmes

Enterprise security teams are already pairing frontier models with large human patch queues. Bank of America's reported Mythos-driven contractor surge is one reminder that finding issues faster can simply move the bottleneck to merge and regression risk. Snyk's pitch is that mergeable, engine-verified fixes shrink that bottleneck; that remains a vendor claim until customers publish outcomes.

Compared with open vulnerability research harnesses, Evo is a commercial platform feature set glued to Snyk's existing scanners, intelligence graph and marketplace posture. Compared with generic LLM 'fix my repo' chats, Snyk's differentiator is the insistence that a deterministic engine, not the proposing model, must confirm the finding and the fix.

Caveats and what to verify yourself

Most of Evo Agentic AppSec is still preview or experimental. Open preview for Remediation Agent does not mean every Snyk plan or every language ecosystem is covered on day one; the launch focus for fixes is Snyk Open Source issues. Experimental agents may never leave select-customer status.

All mergeable-fix and disappearance-rate statistics quoted above come from Snyk. AI Lookout has not re-run VulnBench or VulnBump. Teams evaluating the product should ask for the exact benchmark definitions, language coverage, break-build false-positive rates and how often humans still reject queued fixes.

Finally, blocking packages at install time and auto-opening fix pull requests both change developer trust. Policy design for what an agent may fix without approval is as important as the model that drafts the patch.