Adversarial assessment · code & AI systems

Adversarial intelligence for the systems you ship.

Red Intelligence assesses your code and AI agents against the techniques real adversaries use — grounded in MITRE ATLAS, ATT&CK, and the OWASP LLM Top 10. The analysis runs in your browser, so your code never leaves the page.

Client-side analysis — nothing is uploaded, stored, or logged.
Grounded in
MITRE ATLASMITRE ATT&CKOWASP LLM Top 10CWE
The assessment

Choose what to test for, then add your code.

Ten weakness families, grouped from MITRE ATT&CK, MITRE ATLAS, and the OWASP LLM Top 10. Select one to focus the assessment, or run all ten at once.

Drop a source file, or click to choose
.py · .js · .ts · .java · .go · .rb · .php · .yaml — up to 2 MB
or load a sample vulnerable file
Testing for
Full assessment — all ten classes
Parsed locally in your browser. No network calls, no telemetry.
Assessment
0
—
exposure index / 100
The engine

More than a linter. A red-team analyst.

The in-browser scan catches the obvious patterns in seconds. Connect the deep-scan backend and the same findings run through dataflow analysis, cross-tool corroboration, and an automated verifier that removes false positives before a person ever reads them.

01

Dataflow taint analysis

Traces untrusted input through assignments and calls to dangerous sinks, so an injection is flagged only when attacker data can actually reach it — not on every risky-looking string.

02

Cross-tool corroboration

Bandit, pip-audit, and the taint engine each vote on a finding. Agreement raises confidence and severity; a lone noisy hit is demoted rather than paged out.

03

Variant analysis

Once a real weakness is confirmed, the engine sweeps the codebase for the same shape — the bug class, not just the single line you happened to open.

04

Automated verifier

A deterministic verifier re-checks each candidate against sanitizer and framework rules, labels likely false positives, and ranks what a human should read first.

05

Framework mapping

Every finding carries its CWE and the MITRE ATT&CK or ATLAS technique it enables, so a report lines up with the threat models your team already works from.

06

Authorized by design

Red Intelligence assesses and reports weaknesses. It does not run live exploits against third-party targets — every engagement is scope-gated and non-destructive.

Runtime safety

Agent Tripwire

Static analysis finds weaknesses before you ship. Agent Tripwire watches your AI agents after you ship — a synchronous monitor that inspects each tool call an agent is about to make and can hold, roll back, or kill it before it runs.

Pre-execChecks before execution. Shell, filesystem, and network calls are evaluated against policy before they run — not flagged after the damage is done.
<100msSynchronous control loop. Monitoring runs in-line, so a blocked action is stopped mid-flight rather than noted in a log no one reads.
CanaryTripwire tokens. Honeytokens seeded in context reveal prompt-injection and exfiltration attempts the instant an agent touches them.

From a file to a mapped report in four steps.

First

Pick an attack class

Focus on prompt injection, access control, deserialization — or run all ten.

Second

Add your code

Drop in a file. It's parsed in the browser; nothing is sent to a server.

Third

Read the findings

Each weakness arrives with evidence, severity, and its CWE and MITRE technique.

Fourth

Fix and re-run

Apply the remediation, re-assess, and watch the exposure index fall.

Test your code the way an adversary would.

No account, no upload. Pick a class, add a file, and read the report.

Assess your code