Code analysis pass lands in Shannon 3.0 pentester

Keygraph on September 7 released Shannon 3.0. The main addition is a stage that studies an application’s source code and feeds what it finds to the agents that attack the live app. Keygraph also ran its own comparison, which puts a Shannon scan at between $6.10 and $115 in model costs. The two commercial tools it was set against, Aikido and XBOW, cost $4,000 per scan each.

As Testingcatalog reports, Shannon is an open-source CLI agent for web applications and APIs. It reads the code, works out where the app is likely to break, then tries real exploits against the running target. Anything it can’t demonstrate against the live app is thrown away, a rule the project sums up as “no exploit, no report.” Version 3.0 is available under AGPL-3.0 and runs locally through npx or from source. In the project’s GitHub repository, Keygraph says the exploitation step uses browser automation and command-line tools against the app and its APIs, and a finding only reaches the report if it comes with a working proof of concept.

How the code analysis pass works

The new pipeline has several stages and is adapted from the open-source Mantis security review skills. It runs alongside Shannon’s existing recon and vulnerability agents and starts by building its own model of the application: components, interfaces, dependencies, data flows, trust boundaries and high-value assets. With that map in hand, agents trace attacker-controlled data from the point where it enters the app to the sensitive operation it eventually reaches. This is taint analysis, the static-analysis technique of following untrusted input through code, carried out here by LLM agents.

Candidates from that trace pass through four filters before any exploit is attempted. They are deduplicated, challenged by a separate review agent, checked for whether they could actually work in production, and confirmed against their code paths. Only then do they join the exploitation queue, together with whatever the recon and vulnerability agents found.

The no-exploit rule still applies at the end: a code analysis hypothesis never becomes a finding on its own. Static analysis widens the set of things Shannon tries, and the live app still decides what gets reported. The pass is opt-in through a config flag, so existing setups keep their old behaviour until someone turns it on.

The Photoview comparison

The benchmark builds on someone else’s study. Doyensec, in work sponsored by Aikido, ran Aikido and XBOW against two self-hosted applications. Keygraph took one of those apps, Photoview 2.4.0, set up a matching deployment, ran Shannon 3.0 on it with three different models, and scored the results itself. The figures:

• DeepSeek v4 Flash: 18 findings for $6.10
• Grok 4.6: 10 findings for $35.07
• Opus 5: 24 findings for $115, with one false positive

All three runs flagged the critical pre-auth SQL injection in Photoview, which the project has since patched. The Opus 5 run caught six of the seven issues fixed in that patch round. On its open-source page Keygraph leads with the cheapest result, ”$6.10 in tokens to catch the critical pre-auth SQL injection in Photoview 2.4.0.”

Doyensec’s study, which covered both apps, lists 32 validated findings for Aikido and 7 for XBOW, at $4,000 per scan for each tool. Even the most expensive Shannon run, Opus 5 at $115, cost less than 3% of a single one of those scans. Shannon, though, was run on only one of Doyensec’s two apps, and the people who scored its findings work for the company that makes it. So the cost gap is solid, while the finding counts from the two exercises don’t measure quite the same thing. The spread between models is useful on its own terms: on the same target, Grok 4.6 cost almost six times as much as DeepSeek v4 Flash and reported fewer findings.

Built for the pipeline

The rest of 3.0 is aimed at running Shannon often. Keygraph’s stated case for the project is that teams now ship code constantly with tools like Claude Code and Cursor, while a penetration test happens once a year. Shannon is meant to run against every build or release to close that gap.

To support that, the CLI has been rebuilt. Scan status, phase progress and per-agent logs now show up readably in the terminal. Progress is checkpointed into named workspaces, so an interrupted run picks up where it stopped. On a long scan with an expensive model, that means a crash doesn’t cost the whole token bill again.

An official GitHub Action and a GitLab CI/CD component run the same scan inside a pipeline. They keep reports and logs as build artifacts and can fail the build when Shannon proves a vulnerability above a severity threshold you set. Since a finding exists only once an exploit has worked, a failed build points to a demonstrated problem, not a guess. Reports come out as PDF and Markdown with the evidence attached, plus SARIF 2.1.0, the format GitHub code scanning reads, so findings can show up alongside other code-scanning alerts.

Shannon runs self-hosted, and the operator supplies the model. Supported providers are Anthropic, OpenAI, xAI and AWS Bedrock, plus local runtimes and compatible endpoints such as Ollama, vLLM, LM Studio, OpenRouter and LiteLLM. Orchestration runs on Temporal. That is also why the benchmark is broken down by model: the scan’s price is your provider’s token bill, and the Photoview runs show it varying by almost twentyfold depending on which model you pick.

Keygraph and the paid tier

Keygraph is a San Francisco company founded in 2024 by Varun Sivamani, previously engineering lead for HRIS and payroll at Lattice. It is backed by Pear VC and Authentic Ventures. The project had passed 47,000 GitHub stars before this release. It is named after Claude Shannon, and the company describes pentesting as an information problem in which each probe reduces uncertainty about the state of a system.

The open-source edition costs nothing beyond your model usage. Keygraph also sells a commercial platform that runs an extended build of the same agent continuously. Keygraph Pro is cloud-hosted at $50 per active developer per month, an Enterprise tier runs self-hosted, and a free Community Program covers eligible nonprofits and early-stage startups. The company says the same engine sits under every edition. It pitches Shannon Open Source as a complete pentester in its own lane, aimed at developers and small teams testing staging environments they own.

Keygraph also limits its own claim. It says Shannon is not a replacement for human pentesters. The tool is meant to cover the long tail of internal apps, APIs and fast-moving services that rarely get tested at all, so that expert human time goes to the riskier targets.