Terms of Use

By downloading, installing, or running Bouclier.ai (the "Software"), or by accessing the website at bouclier.ai (the "Site"), you agree to these Terms of Use. If you do not agree, do not install or use the Software, and do not access the Site. These terms apply to all components of the Software including the macOS application, the bundled MCP wrapper, the local proxy, the regex pattern library, the on-device classifiers, the Site, and any related artifacts, documentation, or sample data.

The Software is a research prototype published as the output of independent security and machine-learning research. It is not a commercial product. We do not market, sell, licence for a fee, or commercially distribute the Software, make no representations regarding its fitness for any operational purpose, and offer no service-level agreement, paid support tier, professional services, uptime commitment, or commercial obligation of any kind.

The Software is published under the Apache License, Version 2.0; you are welcome to read, audit, modify, fork, and rebuild the source. The project's intent is research output — independent security testing, academic engagement, and contributor feedback are welcome. Deployment of the Software in any environment where its failure would cause loss, harm, regulatory consequence, or third-party reliance is expressly outside the project's intended use and is undertaken solely at the deploying party's risk.

Pre-1.0 version numbering reflects this status. APIs, detection behaviour, file formats, default settings, the detection pattern set, the refusal thresholds, and the scope of inspected traffic may change without notice between releases. Features described in marketing materials, documentation, the README, the changelog, or in-app text may be removed, altered, or replaced at any time. You must not rely on any specific behaviour persisting across releases.

THE SOFTWARE IS PROVIDED "AS IS" AND "AS AVAILABLE", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, ACCURACY, COMPLETENESS, NON-INFRINGEMENT, OR ANY WARRANTY ARISING OUT OF COURSE OF DEALING OR USAGE OF TRADE.

Without limiting the foregoing, we make no warranty that the Software will detect or block any specific prompt injection, preserve any specific behaviour across sessions or releases, be free of errors, operate without interruption, be secure against any specific threat, interoperate with any third-party tool or service, or meet any regulatory or compliance requirement.

The Software inspects requests for prompt-injection patterns in untrusted content — tool results and other model-visible text your agent fetched rather than text you typed. This detection is probabilistic, pattern-based, and best-effort: it has the false-positive and false-negative characteristics typical of such systems, and a determined attacker can evade it. It is defence-in-depth, not a guarantee, and must not be relied upon as the sole control protecting you from prompt injection or data exfiltration. Specifically and without limitation:

• Scope. Only traffic that is explicitly routed through the Software's gateway (by configuring ANTHROPIC_BASE_URL / OPENAI_BASE_URLor equivalent, or via the Software's own shell/environment wiring) is in scope. Any process, tool, or connection that does not route through the gateway is entirely outside the Software's reach and unaffected by it.
• Evadable. The detector matches known attack families and the cheap obfuscations around them. A payload deliberately shaped to slip past it can and will succeed. A clean pass is not evidence of safety.
• Monitor by default. Out of the box the Software inspects and logs but does not block; it forwards every request. Blocking (refusing a flagged request) is opt-in, per install or via MDM. A monitor-mode install detects but does not stop anything.
• Body size cap. Requests above a configured size threshold are forwarded without inspection, so that large payloads (e.g. file uploads) are not slowed or blocked.

The Software never modifies a request: a flagged request is refused locally with a 422, never rewritten, and your own prompts are always forwarded unchanged. The attachment-PII detection engine described in earlier versions of these Terms is not on any live request path.

Nothing in the Software, the website, the documentation, or related materials constitutes a representation that the Software, alone or in combination with anything else, will cause its user to comply with any law, regulation, contract, or industry standard, including but not limited to the EU General Data Protection Regulation, the Health Insurance Portability and Accountability Act, the Payment Card Industry Data Security Standard, the California Consumer Privacy Act, the EU AI Act, SOC 2, ISO 27001, or any equivalent framework. Compliance with such instruments is the sole responsibility of the user and the user's organisation.

• You are solely responsible for the content of the prompts you send to AI providers and for the use you make of their responses, whether or not the Software was active at the time.
• You are solely responsible for evaluating whether the Software is appropriate for your environment, your data, and your obligations. You must not deploy the Software in any environment in which a detection failure — a missed injection, or a false positive that refuses legitimate work — could cause harm to you, to your employer, to your customers, or to any third party.
• You are solely responsible for the configuration of every Software setting, including MDM-managed values, whether blocking is enabled, and which processes are routed through the gateway. Misconfiguration, running in monitor mode, or simply not routing a given tool through the gateway, may cause an injection to reach your agent uninspected or unblocked.
• You must comply with the terms of service of every AI provider you send traffic to. The Software does not relieve you of obligations arising under your contracts with those providers.

TO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE LAW, IN NO EVENT SHALL BOUCLIER.AI, ITS CONTRIBUTORS, AGENTS, AFFILIATES, OR LICENSORS BE LIABLE FOR ANY INDIRECT, INCIDENTAL, SPECIAL, CONSEQUENTIAL, OR PUNITIVE DAMAGES, OR ANY LOSS OF PROFITS, REVENUE, DATA, GOODWILL, USE, OR OTHER INTANGIBLE LOSSES, ARISING OUT OF OR RELATING TO YOUR USE OF OR INABILITY TO USE THE SOFTWARE, WHETHER BASED ON WARRANTY, CONTRACT, TORT (INCLUDING NEGLIGENCE), PRODUCT LIABILITY, OR ANY OTHER LEGAL THEORY, AND WHETHER OR NOT WE HAVE BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

OUR TOTAL CUMULATIVE LIABILITY ARISING OUT OF OR RELATING TO THE SOFTWARE, REGARDLESS OF THE FORM OF ACTION, SHALL NOT EXCEED FIFTY EUROS (€50) OR THE AMOUNT YOU HAVE PAID TO BOUCLIER.AI FOR THE SOFTWARE IN THE TWELVE (12) MONTHS PRECEDING THE EVENT GIVING RISE TO THE LIABILITY, WHICHEVER IS LOWER.

Some jurisdictions do not allow the exclusion or limitation of certain damages or warranties. In such jurisdictions, the foregoing exclusions and limitations apply to the maximum extent permitted by law.

You agree to indemnify, defend, and hold harmless Bouclier.ai, its contributors, agents, affiliates, and licensors from and against any and all claims, damages, obligations, losses, liabilities, costs, debts, and expenses (including but not limited to legal fees) arising from your use of, or inability to use, the Software; your violation of these Terms; your violation of any third-party right, including without limitation any copyright, property, privacy, or data-protection right; or your violation of any law, rule, or regulation.

The Software is designed to process detection locally on your device. We do not operate servers that receive prompt content, attachment content, audit logs, or user telemetry. The only outbound network calls initiated by the Software are described in the Privacy Policy. You acknowledge that we therefore have no ability to recover, restore, undo, or audit content processed by the Software on your behalf, and that we cannot intervene in any interaction between the Software, your computer, and the third-party AI providers you send traffic to.

The Software re-issues requests you route through its gateway to third-party AI providers. Text prompt bodies and HTTP request headers are forwarded byte-for-byte; the Software never modifies an outbound request — it forwards it unchanged or, when injection is detected and blocking is enabled, refuses it locally as described in Section 4. We have no control over those providers, their terms, their data handling, their availability, or their pricing. Your use of those providers is governed by your agreement with each of them. Content delivered to each provider is governed by that provider's own terms.

The Software incorporates open-source components, including (without limitation) the Meta Llama Prompt Guard 2 integration (governed by the Llama 4 Community License; the model weights are not bundled — the optional on-device ML tier is inactive unless a model is supplied locally), Microsoft Presidio recognition patterns derived from public references, and the Swift, NIO, GRDB, and CryptoKit ecosystems. The corresponding notices are bundled with the Software and reproduced in the project's LICENSE and NOTICE files. These components remain governed by their respective licences.

We may release updates to the Software through Sparkle or other mechanisms. Updates may add, remove, or change behaviour. Continued use of the Software after an update constitutes acceptance of the updated Software and these Terms. We may modify these Terms at any time by publishing a revised version on the website; the "Last updated" date above identifies the current version.

You may stop using the Software at any time by uninstalling it. We may discontinue the Software, suspend distribution, or terminate availability of any feature at any time and for any reason, without notice or liability. Sections 3, 4, 5, 7, 8, 10, 14, and 15 survive any termination.

These Terms are governed by, and shall be construed in accordance with, the substantive laws of Switzerland, to the exclusion of its conflict-of-laws rules and to the exclusion of the United Nations Convention on Contracts for the International Sale of Goods (CISG).

Any dispute, controversy, or claim arising out of, related to, or in connection with these Terms, the Software, the Site, or any matter governed by these Terms — including disputes regarding their existence, validity, breach, termination, or non-contractual obligations connected to them — shall be subject to the exclusive jurisdiction of the ordinary courts of the Canton of Zug, Switzerland.

To the extent that mandatorily-applicable consumer-protection law of your habitual residence affords you the non-waivable right to bring suit before the courts of that jurisdiction, that right is preserved. Nothing in this clause prevents us from seeking injunctive, declaratory, or other equitable relief before any court of competent jurisdiction in respect of a threatened or actual infringement of intellectual property rights, breach of confidentiality, or unauthorised use of the Software.

If any provision of these Terms is found unenforceable, the remaining provisions shall continue in full force, and the unenforceable provision shall be construed to give effect to its intent to the fullest extent permitted by law. These Terms, together with the Privacy Policy and any licences accompanying open-source components, constitute the entire agreement between you and us with respect to the Software, and supersede any prior or contemporaneous communications.

Contact: apps@superstellar.io

If anything in these Terms is unclear, contact us before you install or use the Software. We will not interpret your continued use as acceptance of any clause you have raised in good faith for clarification.