InjectShield

InjectShield is a prompt-injection firewall for AI agents. It is built to stop prompt injection before text reaches an agent, and it scans text such as git commits, web pages, files, emails, user input, and tool output.

The service is delivered as a drop-in REST API. Its documented workflow combines three parts: a heuristic ruleset, a semantic classifier, and context-aware scoring. The ruleset covers about 20 categorized regex patterns for instruction overrides, role hijacks, ChatML and Llama special tokens, exfiltration channels, schema attacks, and invisible-Unicode smuggling. The semantic classifier uses Cloudflare Workers AI and a transformer model to catch paraphrased or obfuscated attacks that the regex layer misses, while capped contribution is used to reduce false-positive flips on benign text. Context-aware scoring assigns different sensitivity to the same text depending on where it appears, and the API supports low, medium, and high sensitivity settings per integration point.

InjectShield provides a live demo, usage snippets for curl, Python, and Node.js, and an API endpoint at api.injectshield.dev/v1/scan. The page also notes that it is built on Cloudflare and uses Cloudflare edge infrastructure, with a stated p99 response time of under 50 milliseconds. It is described as self-serve and gives a free API key by email.

Pricing is listed as free for 10,000 requests per month with no credit card required. Paid tiers include Hobby at $29 per month for 500,000 requests, Team at $99 per month for 5 million requests, and Pro at $499 per month for unlimited requests. The Team tier adds custom patterns, team accounts, and priority support, while the Pro tier includes SLA, no-logging mode, and dedicated support. Overages on metered tiers are $0.50 per additional 1 million requests. The open-source ruleset is available at github.com/bch1212/injectshield.