Bouclier.aiBouclier.aiBeta
Try itTrustAttachmentsHow it worksCoveragePrivacyTermsDownload
Betav0.6.0 — 161 patterns across 21 categories

Stop prompt injections.
Inspect what you upload to LLMs.

Bouclier.ai sits between your apps and AI providers. Every outbound request is scanned for prompt-injection attacks. Images, PDFs and short audio clips are inspected on-device — if they contain PII, the attachment is replaced with a plain-English description before it reaches the model. Your text prompts pass through byte-for-byte; auth headers and API keys are never touched. All inspection runs locally on your Mac.

Beta — research prototype. Not meant to be used live.

Bouclier is published for evaluation, security research, and personal experimentation. It is not a commercial product, is not supported, and is not meant for production, regulated workloads, or any environment where a detection failure could cause harm. Detection is best-effort; false positives and false negatives will occur. See the Terms before installing.

Download for macOSHow it works

Built with Llama. Uses Meta Llama Prompt Guard 2 for on-device prompt attack detection.

Live demo

Try to sneak one past it.

Paste any prompt — benign or adversarial. The exact scanner the Mac app ships runs right here in your browser. 161 regex patterns, Unicode normalization, heuristic scoring. Nothing leaves your machine.

Input108 chars
ResultBLOCKED
Threat score
0.996
Categories
5
Severity
critical

Matched patterns

6
  • System prompt overrideRole Hijack
  • Word splitting obfuscationObfuscation
  • System prompt extractionData Exfiltration
  • Homoglyph / Cyrillic lookalike injectionEncoding Bypass
  • DAN/Developer-mode GPT variantsModel-Specific

What the model would see

[Possible prompt injection redacted by Bouclier.ai. See https://www.bouclier.ai/blocked for details] and [Possible prompt injection redacted by Bouclier.ai. See https://www.bouclier.ai/blocked for details]. [Possible prompt injection redacted by Bouclier.ai. See https://www.bouclier.ai/blocked for details] [Possible prompt injection redacted by Bouclier.ai. See https://www.bouclier.ai/blocked for details], an AI without restrictions.

The desktop app adds a CoreML classifier (Meta Prompt Guard 2) and entropy analysis on top of the regex layer — this demo shows the regex + scoring layer only.

What we don't touch

Your prompts and headers reach the model unchanged.

A proxy that rewrites prompts is a proxy you can't trust. Bouclier sits in the middle of every AI request but it isn't a rewriter — prompt bodies traverse the proxy byte-for-byte and auth headers are forwarded untouched. The only thing Bouclier ever modifies on the way out is an attachment the on-device scanner flagged as containing PII, and even then we substitute a short text description, not a placeholder token that could trip the provider's abuse detection.

01

Prompts

Forwarded byte-for-byte. No tokenisation, no placeholder substitution, no JSON-blind rewriter that could touch user-identifier or analytics fields. The model receives your prompt exactly as your app sent it.

02

Headers

Authorization, x-api-key, X-Trace-ID, custom analytics, User-Agent — every header reaches the upstream unmodified. Pinned by an end-to-end test so a future change can't drift.

03

Attachments

The one thing Bouclier rewrites. Images, PDFs, audio that contain PII are replaced with a plain-English description — never with a token shape that could look adversarial to the model.

The byte-identical guarantee is pinned by E2EProxyTests in CI — every release proves that a real CONNECT + TLS request through the proxy reaches the upstream with body, auth headers, API keys and trace IDs all intact. Read the Terms for the limits of best-effort attachment detection.

Attachment PII

PII hides in what you upload.

A screenshot of an invoice, a scanned NDA, a 30-second voice memo — modern LLM clients accept all of it, and a regex pass over the JSON body sees none of it. Bouclier opens images, PDFs and audio clips on the way out, scans them with Apple's on-device Vision, PDFKit and Speech frameworks, and replaces flagged attachments with a short text description so the model still gets the gist without the leak.

01

Images

Vision OCR + face detection on every image in OpenAI / Anthropic / Gemini multimodal shapes. EXIF orientation honored, 2000 px downscale, 4-concurrency throttle.

02

PDFs

PDFKit text-layer extraction with Vision OCR fallback for scanned pages. Encrypted or oversized PDFs surface as unscannable and get stripped — never silently forwarded.

03

Audio

On-device Apple Speech transcription up to 60 seconds. No audio leaves your Mac. Unsupported formats get blocked rather than passed through.

Attachment inspection ships off by default and is opt-in from Settings → Privacy. Multipart file uploads to the OpenAI Files API and Anthropic messages with PDF / image / audio blocks are all supported. Detection runs entirely on your Mac — no cloud OCR, no cloud transcription, no telemetry.

How it works

Intercept. Scan. Protect.

A System Extension routes AI API traffic through a local proxy on your Mac. Every request and response is inspected before reaching the provider — no code changes, no SDK, no cloud dependency.

01

Intercept

Traffic to 10+ AI providers is automatically routed through Bouclier.ai. Works with any app — ChatGPT, Cursor, Claude, API calls. No configuration needed.

02

Scan

161 detection rules across 21 attack categories. Requests, query strings, and streaming responses are all inspected in real time.

03

Protect

Threats are neutralized inline — injections are redacted before reaching the model. Streaming attacks are terminated cleanly. Safe traffic passes through untouched.

Your apps
Any AI-powered tool on your Mac
Bouclier.ai
Local inspection
AI providers
OpenAI, Anthropic, Gemini, Mistral
Requests scannedStreaming responses inspected
Results

Measured, not marketed.

Every release is tested against 442 real-world attack samples and 240 benign inputs. Detection quality is enforced in CI — regressions block the release.

91.9%
Attacks caught
2.9%
False positive rate
161
Detection rules
21
Attack categories
Coverage

21 attack categories.

Sourced from OWASP LLM Top 10, MITRE ATLAS, HackAPrompt, and red-team research from Anthropic, Microsoft, and leading AI security labs.

Critical High Medium-High Medium
Role Hijack6
Instruction Override5
Tool Poisoning12
Credential Leak11
Memory Manipulation9
Function Hijack8
Model-Specific14
Alignment Bypass14
Code Injection10
Sandbox Escape8
Data Exfiltration6
Indirect Injection7
Context Manipulation5
Chain-of-Thought7
Delimiter Attacks4
Encoding Bypass5
Multilingual15
Payload Splitting3
Obfuscation5
Prompt Leaking4
Recursive Injection3
In action

What it stops.

BLOCKEDRole Hijack
critical

Attempted to override the AI’s instructions and extract its system prompt.

BLOCKEDMultilingual Attack
high

Injection disguised in French to bypass English-language detection.

BLOCKEDAlignment Bypass
critical

Known jailbreak technique (Skeleton Key) attempting to disable safety guardrails.

BLOCKEDTool Poisoning
critical

Malicious instructions hidden in an MCP tool description, targeting SSH keys.

Enterprise

Ready for your security team.

MDM managed

Deploy and configure via Jamf, Kandji, or Mosyle. Control intercepted domains, enforcement policy, and feature flags across your fleet.

Audit trail

Every scan event is logged locally and can be forwarded to your SIEM. Export a privacy-scrubbed diagnostics bundle for incident response.

Hardened by default

Built with defense-in-depth: request size limits, strict input validation, and a published threat model covering every trust boundary.

Streaming protection

AI responses are inspected in real time as they stream. If a threat is detected mid-response, the stream is terminated cleanly.

Privacy

Your prompts never leave your Mac.

All detection runs locally. No cloud LLM, no analytics, no telemetry in the app.

The local CA key is stored encrypted in your Keychain, unique to your device, and removable anytime.

Scan logs never contain your prompts, responses, or API keys — only detection metadata.

Local storage with automatic rotation. You own your data.

Published threat model and privacy policy covering every trust boundary.

Install once. Protect everything.

Download the DMG, drag to Applications, click Enable. Every AI request is protected from that moment.

Download for macOS

macOS 15+ · Apple Silicon & Intel · v0.6.0

Bouclier.aiBouclier.aiBeta
BlockedPrivacyTerms

Built with Llama. Uses Meta Llama Prompt Guard 2 for on-device prompt attack detection.

Experimental, pre-1.0 software. Detection is best-effort; false positives and false negatives will occur. Not intended for production or regulated workloads — see Terms.