Skip to main content

AI security for apps, agents & teams

Detect prompt injection. Redact sensitive data.

Check prompts and AI replies, redact detected secrets and personal data, and hold risky agent actions for review.

The public playground runs on our server and is rate limited. Use sample data, not real secrets.

One text, one decisionRecorded 2026-11-09
Input · marked by the engine

Contact [email protected]1 or +1 415 555 01322.

REDACT

Found: Email address · Phone number

What happens next

Contact [REDACTED_EMAIL] or [REDACTED_PHONE].

Engine reason: Allowed after sensitive values were redacted: Email address, Phone number.

Findings, scores & timings
  1. 1Email addressREDACT · MEDIUM · score 25

    An email address was detected.

  2. 2Phone numberREDACT · MEDIUM · score 25

    A phone-like number was detected.

REDACTrisk 33 / 100

Recorded from the live engine on 2026-11-09. It took 21 ms on the server.

One request, not a performance benchmark. A verdict is one engine's opinion on one text, and it misses things. On 26 blind test cases it caught 16.

Run a live example
Sample text · public, rate-limited engine

A verdict is one engine's opinion on one text, and it misses things. On 26 blind test cases it caught 16.

Paste your own text in the playground.

Choose your starting point

Your workflow. A clear way in.

Build with the API, review team controls, or connect an existing workflow.

  1. For developers

    Check the prompt before your model, and the reply before your user.

    Build with the API
  2. For security teams

    Review the controls for ChatGPT, browser tools and coding assistants.

    Explore team controls
  3. An automated workflow

    Find an integration, check its maturity, and inspect the action before continuing.

    Browse integrations

Evidence, not a guarantee

Every number comes with the test it came from.

The limit sits next to the number, in the same view, so you can decide how much to trust it.

Median 20 ms

Server handling time, 7 live API calls on 2026-11-09 (range 14 to 64 ms)

A sample, not a benchmark. The same calls took a median 344 ms round trip from one machine.

All results, test conditions & limits
Results, the test behind each one, and why not to over-trust it
NumberTestWhy not to over-trust it
100.00% recall2,200 synthetic attacks in the public benchmarkWe wrote the tests and tuned the engine against them.
98.40% recall1,122 attacks, all regression test sets combinedIncludes sets used for tuning.
61.54% recall (about 43 to 78%)26 blind held-out casesA small sample, and the figure closest to real use.
0.00% false positives1,000 benign controls in the public benchmarkSynthetic text written by us. Real users write messier text.
0.00% false positives322 benign controls in the regression runA different, smaller set than the row above. Same limit: synthetic or hand-written text.
p50 9.30 ms, p95 17.83 msAnalyzer only, 3,200 cases, one local CPU run (2026-07-22)Excludes the API, network and authentication. Not run under load.
Median 20 msServer handling time, 7 live API calls on 2026-11-09 (range 14 to 64 ms)A sample, not a benchmark. The same calls took a median 344 ms round trip from one machine.
Read the method and the limits.

Choose a plan

Free to start. Every plan runs the same detection engine.

Higher plans raise limits and add team controls. They do not unlock stronger detection.

See plans

Build with it

From a check to your own code.

Start with the API contract, then choose a framework or integration. The docs explain what to send, what comes back, and when to stop.

One place for three decisions.

  • See

    Prompts are checked for injection, jailbreaks, secrets and personal data before the model reads them.

  • Say

    Replies that carry personal data or secrets are redacted or refused before the user sees them.

  • Do

    Risky tool calls are blocked or sent to a person to approve, and each decision is logged.

Building an AI app or agent? See AI agent security. Setting rules for how your own team uses ChatGPT and coding assistants? See AI user security.

Pick the place your team already works.

Each surface has its own status. What Stable, Beta and Labs mean.

Where it runs, how mature it is, and what it does not do yet
SurfaceStatusWhat it doesKnown gap
APIStableChecks prompts, replies and retrieved text from your server.Needs an API key.
Audit evidenceStableRedacted logs and signed exports.Retention depends on the plan.
IDE GuardBetaFinds secrets and personal data before they reach an AI assistant in VS Code.Detection only on some paths. Other editors not tested.
Browser GuardBetaVisibility and control for AI use in Chrome and Edge.Tested on those two browsers only.
n8n, Make, ZapierBetaGuard steps inside automation workflows.Interfaces may change.
MCP and agent guardLabsRisk review for tool calls before sensitive actions.Experimental. Do not rely on it in production.

Global identifiers and local ones.

Secrets

API keys and tokens, for example AWS keys, GitHub tokens and OpenAI-style keys.

Personal data

Email, phone numbers, card-like numbers, IBAN-like identifiers and SSN-like identifiers.

India

Aadhaar-like numbers, PAN, GSTIN, UPI IDs, IFSC codes and mobile numbers. Indian identifiers and DPDP.

Languages

In a smoke test, one instruction override in each of ten non-English languages was flagged: Spanish, French, German, Portuguese, Russian, Chinese, Japanese, Arabic, Hindi and Hinglish. One phrase per language is not a benchmark.

Known gaps

UK National Insurance numbers are not named. The Aadhaar-like rule can also fire on some card and IBAN digits, and it does not check the number’s checksum yet.

Questions

What is a command layer?

A command layer decides, for every prompt, reply and agent action, whether to allow, redact, block or hold it for approval, and keeps the record.

Does it guarantee security?

No. It lowers risk and it misses things. On 26 blind test cases it caught 16.

How fast is it?

In 7 live calls on 2026-11-09, the server took a median of 20 ms (range 14 to 64 ms), and the same calls took a median of 344 ms round trip from one machine. See the method on the benchmarks page.

Does my text stay private?

The public playground sends your text to our server for analysis and is rate limited. Use sample data. To keep text on your own systems, self-host with Docker, Postgres and Redis.

Which languages and data types does it cover?

It finds email, phone numbers, card-like numbers, IBAN-like identifiers, SSN-like identifiers, common API keys, and Indian identifiers such as PAN and UPI IDs. In a smoke test, one instruction override in each of ten non-English languages was flagged. The known gaps are listed on this page.

Who builds it?

SoterAI is built and run by Yash Chauhan, a solo founder in India.

Test it on text you already have.

Optional analytics

Google Analytics, when configured, and Cloudflare Web Analytics load only if you allow them. Sign-in cookies and product audit records are separate. Privacy policy

Optional analytics is off.