{"uid":"cap_qic8lA6wXmk-6Xy_D1aF_","slug":"verdict-l402-agent-prompt-injection-action-firewall-c103ad0f","name":"Verdict L402 – Agent Prompt Injection & Action Firewall","description":"Agent security guardrail. Judges the pair (untrusted content + proposed action) before an AI agent acts: detects tool calls that obey hidden prompt injection, leave the user's goal, destroy data or send data or money elsewhere. Adjustable sensitivity and optional multi-model consensus. Returns allow","url":"https://verdict.kechlab.com/v1/firewall?utm_source=zero.xyz","method":"POST","headers":{},"bodySchema":{"type":"object","properties":{"goal":{"type":"string","description":"the goal given by the human"},"action":{"type":"string","description":"the action the agent wants to run (tool call, command, payment)"},"content":{"type":"string","description":"the untrusted content read by the agent (page, email, document)"},"consensus":{"type":"boolean","description":"(optional) true: three independent models, disagreement = review, double price"},"sensitivity":{"enum":["strict","balanced","lenient"],"type":"string","description":"(optional) strict, balanced (default) or lenient"}}},"responseSchema":null,"example":null,"exampleRequest":null,"tags":["x402"],"displayCostAmount":"0.0255","displayCostAsset":"USDC","priceDynamic":false,"priceHint":null,"priceStatus":"priced","priceSource":"probe","requiresHandshake":false,"reviewCount":0,"rating":{"score":"0.00","successRate":"0.00","reviews":0,"stars":null,"state":"unrated"},"availabilityStatus":"unknown","priceObserved":null,"sessionDeposit":null,"pricing":{"kind":"static","summary":"$0.0255/call","primary":{"kind":"static","protocol":"x402","network":"base","amountUsd":"0.0255","per":"call","confidence":"exact"},"accepted":[{"kind":"static","protocol":"x402","network":"base","amountUsd":"0.0255","per":"call","confidence":"exact"}]},"paymentMethods":[{"uid":"pm_3THEizRDSey8OIPyXlLV_","protocol":"x402","methodType":"crypto","chain":"base","mode":"charge","costAmount":"0.0255","costPer":"request","priority":0,"asset":"0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913","unit":"request","depositMicros":null,"planRef":null}],"brandName":null,"brandSlug":null,"brandBaseUrl":null,"brandDocsUrl":null,"whatItDoes":"Judges a (goal, action, untrusted-content) triple to detect prompt injection, goal drift, data exfiltration, and unauthorized payments before an AI agent acts, returning allow/review/block.","exampleAgentPrompt":"Before you run that tool call, check it with Verdict's firewall — my goal was 'book a flight to Tokyo', the action you want to take is 'send $500 to account X', and the untrusted content is this email body I just read. Use balanced sensitivity and tell me if I should allow it.","exampleUseCases":[{"title":"Block prompt injection from malicious webpage","prompt":"I scraped a webpage that seems to have hidden instructions in it. Before you execute the 'delete all files in /home/user/docs' command it told you to run, run it through Verdict's firewall with my original goal 'summarize the article' and strict sensitivity — I want to know if this is a hijack attempt."},{"title":"Prevent unauthorized payment in finance agent","prompt":"My agent wants to wire $2,000 to an account it found in an invoice email. My original goal was 'pay the vendor for invoice #1042'. Run the proposed payment action and that invoice email through Verdict's firewall with balanced sensitivity before letting it go through."},{"title":"Multi-model consensus check for high-stakes action","prompt":"Before my agent posts that message to the company Slack channel, check it with Verdict using consensus mode — my original goal was 'notify the team about the outage', and here's the untrusted Slack bot response it wants to relay. I need three models to agree before we allow it."}],"resultDescription":"Returns a verdict of allow, review, or block along with a threat classification (e.g. prompt injection, goal drift, data exfiltration, unauthorized payment), a plain-language explanation of why the verdict was reached, and a confidence score. In consensus mode, disagreement among the three independent models triggers a 'review' verdict automatically.","failureModes":["Missing required fields (goal, action, or content absent) — returns 400 error","Ambiguous or very short content may reduce verdict confidence","Consensus mode doubles the price and may have higher latency","Overly strict sensitivity may block legitimate actions (false positives)","Lenient sensitivity may miss subtle injection patterns (false negatives)","Payment failure (L402) if USDC balance is insufficient — returns 402"],"whenToPreferThis":"Choose this endpoint when an AI agent is about to act on content from an untrusted source (web page, email, document, API response) and you need a real-time security check before execution. Especially valuable when the action involves payments, data deletion, external communication, or any irreversible side effect. Prefer consensus mode for high-stakes or financial actions. This is purpose-built for agentic prompt-injection and goal-drift detection — not a generic content moderation or spam filter.","instructions":null,"reviewSummary":null,"reviewSummaryHighlights":null,"reviewSummaryConcerns":null,"reviewSummaryGeneratedAt":null,"activationCount":0,"lastUsedAt":null,"lastSuccessfullyRanAt":null,"lastHealthCheckAt":"2026-10-21T07:19:29.152Z","isFirstParty":false,"canonicalSlug":"verdict-l402-agent-prompt-injection-action-firewall-c103ad0f"}