"Everyone Sees What Agents Do; We See What They Think"
DETECTION ACCURACY
END-TO-END LATENCY
DETECTION RULES
COMPLIANCE FRAMEWORKS
IN PARTNERSHIP WITH
THE PLATFORM
RUNTIME PROTECTION
Runtime proxy. Every LLM API call routed through the 5-layer detection cascade. One URL change: no code rewrites. Works with OpenAI, Anthropic, Google, Cohere, Mistral, and any OpenAI-compatible endpoint.
PROTOCOL SECURITY
Runtime protection for Model Context Protocol servers. Tool call interception, rug pull detection, and least-privilege enforcement. One line in mcp_config.json: every connected agent protected automatically.
Sits in front of agent traffic and decides every call before it executes. Verifies who is really acting, whether that principal is permitted, and whether the request should be allowed, in single digit milliseconds. Covers legacy agents, third party agents and workloads nobody registered, with no code changes. Built on agentgateway, the Rust proxy stewarded by the Linux Foundation. Self-hosted, air-gapped if required.
Extended AI Detection and Response for multi-agent systems. Every inter-agent message scanned inline before it reaches the LLM. Fleet-level behavioral correlation surfaces coordinated attack chains across agents. Shadow agents discovered automatically. Per-agent trust scores degrade on anomalous behavior. Autonomous quarantine fires when threshold breaks. Microsoft AGT-integrated.
THE FLEET LAYER
A single sensor sees one agent. Fleet Intelligence sees all of them at once. Every xAIDR sensor streams its telemetry to one place, where it becomes a live map of your entire agent fleet: who is running, who is talking to whom, which tools are in play, and where trust is degrading. Correlation across agents catches attack chains that no single sensor could see, a probe on one agent and an exfiltration attempt on another, recognised as one coordinated event. This is the difference between watching agents one at a time and governing them as a fleet.
One live view of every registered agent, its connections, and its trust score
Cross-agent correlation that catches attack chains spanning multiple agents
Fleet-wide trust scoring, with automatic quarantine when standing collapses
Shadow agent discovery, surfacing agents nobody registered
Every signal exportable to your SIEM
Every agent and every principal measured against its own history. Applications behave like metronomes, people are diurnal, scheduled jobs are calendars. Departure from an established pattern is a signal before anything matches a known attack.
WHY DELPHI
The sensor runs inside each agent process. Threats are caught before the LLM call is made, not after the response arrives.
One URL change or one npm install. No new servers, no firewall rules, no changes to your existing LLM provider.
Sentinel Brain correlates behavior across every agent simultaneously. A coordinated attack across three agents becomes one alert, not three separate noise events.
Every agent carries an Ed25519 cryptographic identity via Microsoft AGT integration. Unregistered agents cannot impersonate trusted ones.
Every detection event maps to specific controls across 8 frameworks. OWASP, NIST, EU AI Act, ISO 42001. Evidence generated from runtime. No manual questionnaires.
Define tool whitelists, access scopes, and behavioral guardrails per agent. Policies enforced locally by each sensor. Custom content policies block hate speech, violence, and regulated advice categories.
Delphi consumes identity rather than issuing it. Point the gateway at your existing provider and every token is validated offline against its published keys. No callback into your identity plane, no new dependency in the authentication path, no directory to synchronise.
A principal whose standing collapses is denied on their next request, through every agent they touch, whether or not their token is still valid. Revocation that waits for a token to expire is not revocation.
THE DIFFERENCE
Traditional tools watch for known patterns. Delphi understands intent, correlates behavior across agents, and responds autonomously. Runtime security built for systems that operate without humans in the loop.
Single-agent, input-only scanning
Static rules and policy configs
No visibility into agent-to-agent traffic
Compliance reports built manually
Cannot say which person an agent acted for
Revocation waits for the token to expire
Fleet-level inline scanning across all agents
Intent decomposition and behavioral correlation
Full A2A message inspection and trust scoring
Runtime evidence mapped to 8 frameworks
Verified principal and full delegation chain on every call
Denied on the next request, across every agent
THE DETECTION ENGINE
PRICING
Start free. Scale when you need to.
Five-layer detection cascade: regex, behavioral heuristics, ML, and LLM arbitration
Prompt injection, jailbreak, and multi-turn conversational drift detection
Bidirectional DLP: PII, secrets, and credentials redacted before they leave your system
MCP tool-call interception with OWASP LLM and NIST evidence reports
Usage and cost monitoring with a plain-language AI activity overview: every request, threat, and metric in one live dashboard
Drop-in proxy or SDK, provider-agnostic. Monitor mode by default, safe in production from day one. 100,000 API calls per month.
Custom
Everything in Developer, across your entire agent fleet
xAIDR in-process detection: every inter-agent message scanned inline in under 20ms
Fleet-wide correlation: coordinated multi-agent attack chains surfaced as one alert
Shadow agent discovery with trust-based autonomous quarantine
Compliance evidence across 8 frameworks: OWASP, NIST, EU AI Act, and ISO 42001
Intent-based governance, cryptographic agent identity (Ed25519 and DID), per-agent tool whitelisting. Self-hosted deployment, dedicated CSM, 4hr SLA.
FAQ
One URL change points your existing API calls through the Delphi proxy. No provider changes, no code rewrites. Works with OpenAI, Anthropic, Google, Cohere, Mistral, and any OpenAI-compatible endpoint.
Under 10 minutes for the proxy integration. The npm sensor install takes under 5 minutes. Most teams are running in production the same day.
xAIDR is Delphi's fleet security layer for multi-agent systems. If you run a single LLM application, AI Sentinel is sufficient. If your agents communicate with other agents, xAIDR adds inline A2A scanning, trust scoring, and shadow agent discovery.
OWASP LLM Top 10, OWASP Agentic Top 10, OWASP MCP Top 10, NIST AI RMF, EU AI Act, ISO 42001, AIUC-1, and AWARE. Evidence is generated automatically from runtime detection data. No manual questionnaires.
Yes. 14 days, full access, no credit card required. The free tier includes 15,000 API calls per month with no time limit after the trial ends.
RESEARCH
Technical research, benchmarks, and product updates from the Delphi Security team.
WHITEPAPER
Whitepaper introducing xAIDR, a new category of AI security. 94.5% accuracy with 98.4% precision on a 500-prompt agent-to-agent test suite.
BENCHMARK
How Delphi AI Sentinel achieved 94.5% accuracy and 98.4% precision on a 500-prompt adversarial test suite targeting agentic AI tool calls, MCP poisoning, and cross-agent attacks.
THE PYTHIA CHALLENGE
10 levels. Escalating difficulty. Our proprietary adversarial challenge lets you test the limits of the detection cascade directly. Beat all 10 levels and we will have a conversation.
FROM THE FOUNDER
I spent over a decade in security, breaking into systems as a penetration tester and defending industrial environments where a wrong action has physical consequences. When AI agents started taking real actions, calling tools, moving data, invoking other agents, I saw the same gap I had seen everywhere else: everyone was watching what the model said, and nobody was watching what the agent did. Delphi exists to close that gap. Security that runs where the action happens, so an agent can be stopped before it acts, not audited after.
Anirudh Kotaru, Founder
