Not long ago, finding a subtle but critical flaw in an unfamiliar codebase could take an experienced pentester days: reading the code, tracing how its pieces connect, forming a hypothesis, testing it. Frontier models can now do that same reasoning in minutes, on any codebase. Attackers now have that capability for the first time, at a speed and scale no human team can match.
AppSec teams can’t outrun these attackers using traditional pre-AI tools, nor by relying on humans reviewing all of the code being generated. To give AppSec teams a step up in defensive capabilities, today we’re launching nine Agentic Workflows that find the subtle authentication, injection, and business logic flaws that matter most. Agentic Workflows enables continuous hunting across your code, and builds toward a security program that runs itself.
What’s shipping
Starting today, Semgrep offers nine pre-built Agentic Workflows (in public beta) covering 70+ CWEs:
| | | Traces user-controlled data into dynamic queries that lack effective parameter binding or allow-list validation. | | Traces user-controlled input to unsafe HTML rendering sinks, across reflected, stored, and DOM-based cases. | | Analyzes attacker control, validation, and bypass paths using repository-specific rules. | | Traces untrusted input to shell and process-execution APIs across supported languages. | | Traces untrusted input into query and scripting operations in MongoDB, Redis, and Elasticsearch. | | Flags security-sensitive use of MD2, MD4, MD5, and SHA-1, while filtering out benign checksum, cache-key, and identifier uses. | | Audits application code across web, API, cryptography, access control, and AI security, then filters by confidence and exploitability. | | Audits REST, GraphQL, gRPC, SOAP, JSON-RPC, and webhook APIs, then prioritizes findings by exploitability. | | Audits cryptography, password hashing, key management, signatures, randomness, and TLS, then prioritizes by exploitability. |
Adding to everything Semgrep and the Semgrep Pro Engine already detect, Agentic Workflows run on your existing pool of AI credits. Turn them on for yourself today.
What early access looked like
We ran Agentic Workflows with a small set of design partners to gather feedback before launching our public beta and saw two kinds of reactions. One set of partners manually validated every finding a Workflow surfaced and confirmed virtually all were true positives. Another ran Agentic Workflows results side by side with their own internal review set, expecting to find gaps, and came away impressed by the near-perfect match.
Why we built this
Attackers can now get pentester-grade reasoning on demand, at a level of capability they’ve never had before. AppSec teams need the same capabilities running against their own code, to catch flaws like broken authorization, missing checks, and injection through an unfamiliar path before an attacker does.
That need compounds because so much of what’s moving through review now is AI-generated in the first place. In the past, human review of critical repos and potential exploits could keep up with the pace of development. Now, AppSec teams need to accelerate that review, and apply similar tools that are accelerating the capabilities of attackers.
How it works
Each Agentic Workflow draws on the same toolkit: the Semgrep Pro Engine and Pro rules for deterministic analysis such as taint tracing and interfile analysis, Semgrep Mandoline (a code slicer internal to Agentic Workflows built specifically for handing agents exactly the context and code they need), and frontier AI models such as Claude Opus for reasoning about exploitability.
This is the same approach behind Semgrep AI-powered detection, which already combines AI reasoning with program analysis to find vulnerabilities, like broken authorization, that don’t announce themselves in a simple data-flow pattern. In our own benchmarking, that combination finds 3.5x more true positives than Opus 4.8 running alone, at 19% lower cost per true positive.
Agentic Workflows detection also runs on the same platform we introduced this spring with Custom Agentic Workflows: a Python SDK, a library of Semgrep and AI tools, a CLI for local development, and managed infrastructure that handles orchestration, retries, and cost tracking at scale. These nine pre-built Agentic Workflows are ready to run out of the box. If you want to go further, you can build and customize your own Agentic Workflows on that same foundation.
What’s next
Detection is one part of a longer loop. We’re already seeing strong early results to further augment and automate AppSec processes, by apply Agentic Workflows for triage, remediation, and proactive security hardening. We plan to continue investing in triage and remediation Workflows, and in the context that makes every Workflow more accurate, over time.
Get started
Agentic Workflows is available today, in public beta, to every Semgrep customer. Log in or sign up to try it out.