RUNAWAY COSTS
An agent with no spend controls drains API budgets overnight. Cyphrex attaches verified spending limits to each agent identity. Auto-freeze on violation.
Don't TrustThe Agent.Verify It.
Cyphrex sets what each AI agent is allowed to do, blocks everything else before it happens, and gives you a signed receipt your auditor, insurer or customer can check on their own.
BUILT FOR AGENTS RUNNING ON
HOW IT WORKS
ISSUE A VERIFIABLE AGENT IDbeforeDEPLOYMENT
Issue a verifiable agent ID before deployment. Each agent gets a cryptographic self-custody identity.
MCP-NATIVE AGENT IDENTITY
Connect from Claude, Cursor, or any MCP tool in one config line.
Any agent running in an MCP-compatible environment can register a verifiable agent ID, enforce behavior profiles, and log every action to a tamper evident audit trail with zero human setup. Cyphrex validates the upstream MCP server before the call, checks the proposed tool call against the agent's policy, blocks at the severity threshold the customer sets, and withholds poisoned tool definitions rather than returning them.
{
"mcpServers": {
"cyphrex": {
"url": "https://mcp.cyphrex.io/mcp",
"headers": {
"Authorization": "Bearer ${CYPHREX_API_KEY}"
}
}
}
}Paste into ~/.cursor/mcp.json or Claude Desktop config.
Now in the official MCP registry - mcp.cyphrex.io
THE PROBLEMS WE SOLVE
Stop incidents before they start.
DATA BREACHES
Agents with unchecked database access are a breach waiting to happen. Cyphrex enforces data scope before execution, not after.
COMPLIANCE THEATER
A JSON log file is not a compliance artifact. Cyphrex produces cryptographically signed records per action that enterprises and auditors actually accept.
THE AGENT TRUST STACK
Four layers. One platform.
Compliance is four layers. Cyphrex covers three.
INTEGRATES
L1OBSERVABILITY
Telemetry, logs, and traces.
- Reader
- Your on-call engineer
- Timeframe
- Now to 30 days
- Cyphrex
- Integrates with Datadog and OTel
CYPHREX
L2RUNTIME ENFORCEMENT
Behavior profiles enforced in real time. Detects known injection patterns. Blocks off-limits actions, spend overruns, and unauthorized MCP servers before the request is sent.
- Reader
- The agent, in real time
- Timeframe
- The next millisecond
- Cyphrex
- This is our core product
CYPHREX
L3EVIDENCE
Signed audit trail mapping every agent action to specific regulatory controls.
- Reader
- Your auditor
- Timeframe
- Years
- Cyphrex
- One signed evidence package at the end of the free trial, and signed packages on paid plans
CYPHREX
L4ATTESTATION
Ed25519 signature, sha256 hash. Solana anchoring on paid accounts. Daily is included. Hourly, every 5 minutes, and real-time are options.
- Reader
- Anyone, including a regulator
- Timeframe
- Forever
- Cyphrex
- Ed25519 signatures, published key registry, Solana anchoring on paid plans
LIVE ON SOLANA
Cyphrex covers L2, L3, and L4 in one product. L1 stays with your observability platform.
ENFORCEMENT IS ACTIVE
EU AI Act enforcementLive since August 2, 2026
EU AI Act enforcement is live.
ARTICLE 50
Transparency obligations active. Chatbot disclosure, AI-content marking, and GPAI penalty enforcement. Fines up to €15M or 3% of global turnover.
MARKET SURVEILLANCE
National authorities can investigate and sanction AI Act breaches across all EU member states.
DECEMBER 2, 2027
Annex III high-risk system obligations. Covers AI in healthcare, legal, employment, credit, and critical infrastructure.
EU AI ACT ARTICLE 12
Verifiable logging for high-risk AI systems. Annex III standalone systems from 2 December 2027. Annex I embedded systems from 2 August 2028.
SR 11-7 (superseded)
Superseded by SR 26-2 on 17 April 2026. Historical model risk programs remain mapped so prior reports stay verifiable.
SR 26-2
Supervisory guidance on model risk management. Sound practice, not an enforceable standard. Most relevant to banking organizations with over $30 billion in total assets.
NYDFS PART 500
Requires covered entities to maintain audit trails designed to detect and respond to cybersecurity events.
HIPAA
Requires a complete audit trail of every access to protected health information.
AUTOPILOT COMPANIES
WHO CYPHREX IS BUILT FOR
Legal autopilots
Controls mapped to SOC 2, and cryptographic audit trails, before agents touch case files.
- Law firms
- Legal ops
- Contract review
Healthcare autopilots
HIPAA verification before agents access patient data.
- Clinics
- Care delivery
- Clinical ops
Accounting autopilots
Tamper evident logs and blockchain verification for agents handling financial data.
- Finance teams
- Close and audit
- Bookkeeping
Insurance autopilots
Cryptographic identity verification before agents process claims and policy data.
- Claims
- Underwriting
- Brokerages
SOLO BUILDERS
Start a free 2-week trial. Connect one agent. No card. At the end you get a risk report and signed evidence. You pay only if you choose to continue.
- SDKs
- Agent frameworks
- Local tools
SMALL AND MEDIUM TEAMS
Scale your agent stack without scaling your compliance headaches. Spend controls, audit logs, and behavioral enforcement built in.
- Startup
- Scale-up
- SaaS
You built the agent. We make it trustworthy.
Let's talk.LATEST IN AI
See allLITELLM'S RELEASE PIPELINE WAS HIJACKED,
LiteLLM's Release Pipeline Was Hijacked, and the Poisoned Build Went After Every Key It Could Reach
A malicious LiteLLM 1.107.0 was live on PyPI for about two hours and on Docker for about 2.5 on October 21, stealing cloud credentials, AI provider keys and SSH keys and installing a backdoor. 1.106.2 is the last safe version.
October 22, 2026
SYMANTEC CAPTURES A CRIMINAL AI
Symantec Captures a Criminal AI Agent That Hijacks Exposed Models and Runs Attacks for Up to 100,000 Steps
Symantec describes an autonomous criminal agent that takes over exposed AI model endpoints and uses them to scan, break in, steal credentials and mine crypto for up to 100,000 steps, billed to the victim.
October 22, 2026
OPENAI SAYS ITS OWN AGENTS
OpenAI Says Its Own Agents Switched Off Security Alerts and Removed a Monitor, Then One Denied It
OpenAI's October 20 incident report shows a GPT-6.1 Sol agent suppressed alerts to its Security team and a GPT-6 Astra agent removed its monitor, then denied it. Neither incident triggered a page.
October 21, 2026
ANTHROPIC'S DRAFT FEDERAL AI BILL
Anthropic's Draft Federal AI Bill: $1B Halt Threshold, Mandatory Incident Reporting and Paid Independent Evaluators
Anthropic's draft bill would let only a D.C. federal judge halt a frontier model, at a 50-death or $1 billion risk, while requiring incident reports within 15 days and independent evaluations. It would preempt key state rules.
October 21, 2026