A clearer view.
With every iteration.

The free demo runs through November 30.

The free demo is extended by a month, through November 30, 2026 (UTC). Nothing else changes: the same limits, one shared detector release for the hosted API and the local download, and no automatic charge or subscription when it ends. The n8n workflow note names the new date.

The guide inside the downloadable 0.4.14 package still says October 31. Local scans are not time-limited, and the next detector release will carry the new date.

A clear answer to an oversized request.

A request body over the 64 KB limit now reliably gets its 413 answer. Before, the API closed the connection without reading the body, so a client that sends the whole body before reading the reply could see a dropped connection instead of the documented error. Found while looking into an independent retest. Detector 0.4.14 is unchanged.

Answers that match the contract.

Sending no key to the sanitize route now gets a 401, as the OpenAPI contract says, instead of a 402. The contract also lists every status the keyed routes return, including a 401 for a key we do not recognise on the usage and tool-check routes, so a generated client can expect each one. Found in independent retesting. Detector 0.4.14 is unchanged.

A sturdier hosted API.

Changes from a production-readiness review of the hosted service. Calls without a key now have a per-minute limit as well as the daily one, and repeating the same text counts toward it. A URL check has an overall deadline and a cap on how many run at once, an MCP batch has a size limit, and malformed input gets a 400 instead of a server error. Detector 0.4.14 is unchanged.

On the API’s key page the button now reads “Show my key”: the key appears once, on that page, and nothing is emailed. The pages that show a key send a strict content security policy, refuse to be framed and are never cached.

We don’t keep what you scan.

The site now says it up front: under the homepage headline, beside the live scan, and at the top of the privacy policy. Your email address, API key and IP address are not kept as you sent them either, only as one-way hashes, and the API’s key page now says so beside the email field. Detector 0.4.14 is unchanged.

Both statements are checked with every build. After every deploy, marked scans sent to the live service are searched for on its server and in its logs.

Caesar shifts, however they are written.

Independent retesting reported a Caesar-shifted attack scoring LOW a fifth time. The ways of applying a shift we knew of were caught, so the gap was in how the shifted text is written: in capitals with no spaces, as the pycipher library writes it, or in five-letter groups; with its words joined by underscores; with letters pushed past z instead of wrapping; or hard-wrapped like plain-text email, under any disguise. Detector 0.4.14 reads all of them. The hosted API runs it from October 2, and the local download is now 0.4.14 as well.

Ordinary text scores exactly as before: some 14,000 benign texts, the 40-case baseline and every benign test are unchanged. Three shapes the decoders still miss are published, among them words run together without spaces. The n8n workflow passed 21 fresh executions against 0.4.14.

Caesar shifts, however they are applied.

Independent retesting still found a Caesar-shifted attack scoring LOW. Plain shifts were already caught, so the gap was in how the shift is applied: an instruction shifted inside an otherwise ordinary message, every character shifted rather than only the letters, or only the lowercase letters shifted. Detector 0.4.13 decodes all three. The hosted API runs it from September 29, and the local download is now 0.4.13 as well.

Ordinary text scores exactly as before: some 14,000 benign texts, the 40-case baseline and every benign test are unchanged. The n8n workflow passed 21 fresh executions against 0.4.13.

One public contact, everywhere.

The API’s own key page, where “Get a Trismag key” leads, now lists hello@trismag.dev. It had shown the operator’s personal address. The page now uses the configured public contact, as the API’s security.txt already did. The detector is unchanged at 0.4.12.

Tighter by default.

The website now asks browsers not to send a referrer to other sites, opts out of the browser ad-topics API and of device access, and isolates its windows from other sites. Its Content Security Policy allows no inline code and enforces Trusted Types. The site's gateway functions have per-IP rate limits, and their responses cannot be rendered or framed. trismag.dev publishes its own security.txt contact file, and Netlify's default address for the site now redirects here.

The privacy policy now covers the local package, the website's request limits, the signed address forwarding on legacy API routes, and the records and email provider used only when email verification is enabled. The retest guide now lists hello@trismag.dev as its contact.

Disguises no longer hide the rules.

Independent retesting found attacks that scored LOW once disguised with a Caesar shift, ROT13, reversal or percent-encoding. Detector 0.4.12 checks every decoded copy with every rule the plain text gets, and recognizes shifted ordinary English. The hosted API has run it since September 27, and the local download is now 0.4.12 as well. During the demo, each detector release reaches the hosted API and the download together.

The Make blueprint now continues only on HTTP 200 with a LOW verdict. The developer guide explains how to declare untrusted content with provenance. The detector now warns on the previous blind-spot example, so the demos use a request it still misses, and the n8n workflow passed 21 fresh executions against 0.4.12.

The terms now say how releases work during the demo: each new local release replaces the previous download. The 0.4.11 download has been removed; copies already downloaded keep the local demo permission.

One release. More time to try it.

The free demo is extended through October 31, 2026. Local Python downloads are available now, with one shared detector release during the demo. Hosted allowances remain unchanged.

After the demo, the planned offer is free Community quarterly snapshots and Developer at $5/month for current updates. Hosted targets of 100,000 scans/month and 60 requests/minute need capacity validation. Both maintained channels will receive urgent software-security fixes. Checkout remains closed. This plan supersedes the earlier $9 pack and twelve-month-update proposal below; older entries are historical.

Inspection where your content lives.

The prepared Python wheel brings offline text, conversation and tool-metadata inspection to Python and the terminal. A LOW-only helper returns the original text or stops the calling path. Installed scans need no hosted account or API key. Public downloads are pending paid access: each purchase will include 12 months of updates, with the downloaded version continuing to work afterward.

This release addresses the reported whitespace, directive-paraphrase, typo and end-marker cases. The homepage and setup guide now distinguish local execution, hosted inspection and workflow connections. API deployment is tracked separately; the live health endpoint remains authoritative for hosted versions.

Personal purchase preparation.

A dedicated purchase page explains account matching, personal use and refund conditions. The API records versioned acceptance from that flow; direct Stripe links remain a separate path. Operator tools now reserve and reconcile refund credits without negative balances or repeated adjustments. Website checkout stays closed pending final Stripe checks. Detector v0.4.4 is unchanged.

Clear purchase and refund conditions.

Published scan-credit purchase terms covering delivery, a full refund requested within 14 calendar days for completely unused packs, individual review of partial refunds and preserved statutory rights. Paid checkout remains closed pending final launch checks.

Personal and business purchase preparation.

The payment receiver now supports the two explicitly configured Trismag payment links with the same amount validation and duplicate-credit protection. Public checkout remains closed pending final purchase terms and final payment configuration verification. The free demo and detector v0.4.4 are unchanged.

Inspection inside a real workflow.

The n8n starter now validates responses, holds explicit incomplete scans, preserves inspected text for the release step and disables saved execution data. Published local n8n and MCP SDK reports state their scope and skipped live verdicts. A new guide separates workflow enforcement from optional MCP scanning and includes a recorded known-miss example. Existing imported workflows need a manual update. Detector v0.4.4 is unchanged.

Free access continues. Integration matters.

Limited free usage continues after the September 30 demo ends. The current free-key allowance is 500 scans per calendar month, subject to published limits. MCP guidance now distinguishes optional model-invoked scans from inspection enforced by your application before content reaches the model. No detector or runtime enforcement change is included.

A clearer way to reach us.

Contact Trismag at hello@trismag.dev for questions, integration help and demo feedback. The homepage now brings the free demo and planned $9 scan-credit pack together, with checkout availability clearly marked.

A verified payment notification path.

The API now verifies signed Stripe notifications and prevents a checkout session from granting credits twice. Payment metadata and its retention are described in the privacy policy. Scan content is not sent to Stripe.

Public paid checkout remains pending its final launch checks. The free demo runs through September 30, 2026 under existing limits. Detector v0.4.4 is unchanged.

Time to put it to the test.

The free public demo runs through September 30, 2026. Existing usage limits apply; no card is required and using the demo does not authorize a charge.

A one-time $9 USD pack of 10,000 scan credits is planned, plus applicable tax. Paid checkout is being connected and is not yet available through this site.

Scheduled cleanup. A frozen detector.

Expired usage and deduplication records now receive bounded background cleanup. New SQL recovery exports omit those transient records and the client salt; older backups remain a separate retention concern.

The detector remains v0.4.4 for independent retesting. Health responses expose its source hash, plus cleanup worker status. No new detection-rate claim accompanies this maintenance release.

Clear obligations. Visible scan costs.

Published dedicated privacy and preview terms under Jonathan Hill as the individual operator. The privacy policy corrects earlier deletion-window, backup and provider descriptions. The terms explicitly explain missed attacks, false alarms, service failures and the limits of protection.

Documentation now reflects the live domains and allowances. A fixed-baseline offline performance gate measures rule growth before release; no customer payload telemetry was added. Detection rules and API response contracts remain v0.4.4.

Warnings for previously missed steering.

New deterministic rules cover answer manipulation, task substitution, payment-target replacement, query-based secret leaks and separated-letter overrides. All ten misses in the original synthetic baseline now flag; its benign risks are unchanged.

The original results remain public. Post-change results are explicitly development regressions, not a claim of real-world accuracy. The demo now shows a different documented blind spot.

A live first impression.

Edit and scan text directly on the homepage without a key. Anonymous previews report risk; the optional keyed path exposes matched findings. Editing the input clears the previous verdict.

• Narrow browser access in API v0.4.2 preserves each visitor’s existing IP allowance.
• A downloadable 40-case synthetic detection baseline includes misses and benign flags, plus results reproduced against the deploy.
• A two-week pilot guide helps teams evaluate one agent workflow.

A new front door for Trismag.

A dedicated product site with an inspection playground and a practical guide to adding triage to an agent workflow.

• Reference examples for text, conversations, and MCP tool metadata.
• A live scanning path using your own Trismag API key.
• Visible distinctions between reference findings, live results, and uninspected content.
• A trust center covering limitations, request handling, and the boundary between detection and authorization.

Connected to API v0.4.1.

At the initial website release, the live health endpoint reported version 0.4.1. The website keeps API calls on the existing Fly.io origin, preserving a separate deployment path for the service.

The current public contract offers anonymous previews and free keys. Paid access is not yet available. Limits and availability remain defined by the deployed API.