IN PARTNERSHIP WITH
Inside the agent.
xaidr sensor · Python
The open-source xAIDR sensor runs in the agent’s own process and decides locally, before a tool call executes.
In front of the traffic.
agentgateway · ext_authz
The gateway checks every model, tool and agent-to-agent call against your policy and verifies who is acting.
Inside your walls.
Delphi Core · your cloud or on-premises
Self-hosted in your cloud or on-premises, air-gappable, with no content leaving your network by default.
Self-hosted
Air-gappable
Zero telemetry by default
Defence in depth & Zero trust
Most tools judge what the model said. We judge what the agent is about to do. Denied by default
Text
What was said?
Catches injected instructions, jailbreaks and secrets in prompts, responses and fetched content.
“ignore previous instructions”
Action
What does it do?
Reads the tool call itself, a shell command, a query or an email, and how much harm it can cause.
rm -rf ~/projects, classed as destructive
Virtual patching
Is this a known attack?
Signed patches stop known flaws in agent tools before the vendor ships a fix.
tool-response injection in an MCP server
Behaviour
Is this normal?
Learns how each agent and person works and flags what falls outside it.
first call to payout-update.net
Intent
Was it asked for?
Traces every action back to the request that authorized it.
recipient not in the request
Five checks on every action. Any one can stop it.
An authorized action must trace to an authorized intent.
The console
One console. Every decision, explained.
Compliance
Evidence your auditor can follow.
Every decision maps to the control it evidences.
Each control states its evidence test in plain words.
A control with no runtime evidence is marked Attestation, never Evidenced.
Export an auditor pack: a PDF summary plus one CSV per framework.
EU AI Act
NIST AI RMF
ISO/IEC 42001
OWASP LLM Top 10
OWASP Agentic Top 10
OWASP MCP Top 10
AIUC-1
WHY DELPHI
Security that works at machine speed.
Inline. Not proxied.
The sensor runs inside each agent process. Threats are caught before the LLM call is made, not after the response arrives.
Zero infrastructure changes.
One line in your agent, or one route through the gateway.
Fleet-aware governance.
Delphi Core correlates behaviour across agents. A coordinated attack across three agents becomes one alert, not three separate noise events.
Cryptographic agent identity.
Every agent carries an Ed25519 cryptographic identity via Microsoft AGT integration. Unregistered agents cannot impersonate trusted ones.
Compliance evidence, automated.
Every detection event maps to specific controls across 7 frameworks. OWASP, NIST, EU AI Act, ISO 42001. Evidence generated from runtime. No manual questionnaires.
Governance policies, per agent.
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.
Identity you already own
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.
Data sovereignty by design
Delphi runs entirely inside your perimeter. Self-hosted, air-gapped if you need it, with no backend and no callout in the scan path. Telemetry carries a hash and a length, not your content. Your data, and your sovereignty over it, stay yours.
THE DIFFERENCE
Not a guardrail. Not a policy engine.
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.
Traditional AI security tools
Judge what the model said
Call a cloud API on every request
Rely on detection that adaptive attacks learn to evade
See one app or one gateway
Can’t identify the acting person
Revocation waits for token expiry
Compliance reports assembled by hand
Delphi Security
Judges what the agent is about to do
Decides inside the agent and inside your network
Enforces deterministically on actions
Covers coding assistants, AI apps, agents and agent-to-agent traffic
Verified principal and delegation chain
Denied on the next request, across every agent
Evidence mapped to each control, ready to export
PRICING
Runtime security for every stage.
Start free. Scale when you need to.
Open source
The xAIDR sensor, Apache 2.0.
Free
In-process detection inside your Python agents
Decisions made locally, with no backend required
Rules for prompt injection, tool misuse and agent-to-agent attacks
Runs anywhere you can pip install
Enterprise
The full platform, deployed in your environment.
Talk to us
Delphi Core decision engine and gateway
AI apps and coding assistant governance
Agent builder
Data protection, behaviour analytics and virtual patching
Compliance evidence for seven frameworks
Self-hosted or air-gapped deployment
Dedicated support
RESEARCH
From the lab.
Technical research, benchmarks, and product updates from the Delphi Security team.
WHITEPAPER
xAIDR: Extended AI Detection and Response for Multi-Agent Runtime Security
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
94.5% Agent-to-Agent Accuracy on 500 Adversarial Attacks
94.5% detection accuracy and 98.4% precision across 500 adversarial agent-to-agent scenarios. One of the first public benchmarks for agent-to-agent attacks, as far as we know.
CHALLENGE
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.