The agent identifies entry points: API routes, user inputs, authentication boundaries, trust transitions. From these it builds a map of where untrusted data enters the system.

Trace data flows

From each entry point, the agent traces how data flows through the codebase: across function calls, through middleware layers, into databases and external services.

Identify security invariants

The agent reasons about what must be true for the system to be secure: authorization checks before data access, input validation before use, proper scoping of credentials. Then it looks for where those invariants break.

Validate exploitability

Candidate findings are assessed for reachability before they are reported. Where possible, the agent executes code in the sandbox to confirm the vulnerability is exploitable in practice.

Recommend structural defenses

Beyond patching the specific bug, Tachyon recommends defense-in-depth improvements where they apply: fail-closed defaults, scoped access, rate limiting.

What these findings have in common

Four classes that recur across the advisories above. Each card links to a published example.

Authorization logic flaws

Permission checks enforced on one API surface and skipped on another. In MLflow, unprotected Flask routes and GraphQL endpoints let authenticated non-admin users access and modify restricted resources.

CVE-2025-14297MLflow

Multi-step exploit chains

Sequences of individually-correct operations that combine to break a security boundary. In authentik, a view-only permission on policies reached a test endpoint that executes arbitrary code in the container.

CVE-2026-25227authentik

Indirect data flows

Untrusted input that reaches a dangerous operation through layers of indirection. In Gradio, a proxy_url field from a remote Space's config was added to an allowlist without validation, then used for outgoing HTTP requests.

CVE-2026-28416Gradio

Incomplete mitigations

URL and path validation that stops the obvious vector but not its variants. Indico fetched user-provided URLs in several places without blocking localhost or cloud metadata endpoints.

CVE-2026-25738Indico

Writeups

What happened after we pushed our .env to a public repo

Rahul Govind·July 20, 2026·5 min read

SecurityHoneypotsResearch

Raise the Floor

You can't finish securing yourself. You can finish finding out.

Rahul Govind·May 30, 2026·9 min read

Security

Trivial To Introduce, Impossible to Fix: Why SSRFs are the Trickiest Security Issue in Modern Web Apps

One line of code introduces an SSRF. Fixing it correctly requires aligning URL parsing, DNS resolution, redirect handling, HTTP client behavior, and network policy—all at once, without missing a single edge case.

Rahul Govind·February 27, 2026·8 min read

SecurityWeb Security

Sandboxes Won't Save You From OpenClaw

AI agent misbehavior isn't a sandbox problem—it's a permissions problem.

Aakash Japi·February 24, 2026·5 min read

SecurityAI

CVE-2025-14297: MLflow Authorization Bypass

How Tachyon's autonomous security researcher found an authorization bypass in the open-source MLflow tracking server by reasoning across protocols and surfaces—and why this class of bug is so hard to catch.

Aakash Japi·February 3, 2026·10 min read

Security

Run this analysis on every pull request

The methodology on this page is the same one Tachyon runs against your repositories.

Get started