Agentic Security — Runtime Verification

Agentic Security — Runtime Verification

Runtime verification for Cursor, Claude Desktop, Claude Code, Codex, OpenClaw, and Hermes

Your developers and contractors run Cursor, Claude Desktop, Claude Code, Codex, OpenClaw and Hermes on their machines and servers, with access to files, shells, credentials and MCP tools. EDAMAME, the AI agent trust layer, watches each agent from outside the agent, with no plugin or SDK inside it, and turns what it sees into evidence and access decisions.

SEEN FROM OUTSIDE THE AGENTSimulation · on task
Claude Code
Session in ~/src/payments-api
Task:fix the failing test
EVENTS FROM THE OSprocesses · files · network
new session · Claude Code
10:42:28claudeopentest/ledger.spec.ts
10:42:28claudeexecnpm test
10:42:29npmconnectregistry.npmjs.org:443
10:42:30nodeopensrc/ledger.ts
10:42:30claudewritesrc/ledger.ts
10:42:31claudeexecgit fetch
DIVERGENCEvs the declared task
Low
ATTACK PATTERNSon the device
No match
EDAMAME HUB · POLICY
AI allowlist passes
Access kept
Nothing installed into the agentDetectors run on the device
Intent divergence

Judge each agent by what it did, not by what it said.

EDAMAME compares what an agent was asked to do with what the machine actually saw, from outside the agent.

EDAMAME Security, Agents > Divergence: the divergence monitor maps an agent's processes and traffic against its expected behavior
EDAMAME Security · Agents › Divergence (demo data)

Reads the declared task

EDAMAME reads each agent's transcripts on the host: the task it was given and the tools it called.

Watches the machine, not the chat

Processes, files and network, read from the operating system, outside the agent's own process.

Scores the gap

A divergence verdict with its evidence, and no plugin to install inside the agent.

Attack-pattern findings

Caught by technique, not by indicator.

The same host telemetry catches compromised behavior: what code opens, spawns and sends.

EDAMAME Security, Agents > Exposure: the most recent alertable findings, such as a process reading several credential stores
EDAMAME Security · Agents › Exposure, findings (demo data)

Eleven deterministic checks

Such as token exfiltration, credential harvest, sensitive-material egress, sandbox escape and file tampering, plus an agent turning its own sandbox off or re-spelling a denied command.

Release-gated

Each release must pass 23 attack scenarios and a divergence scenario on four platforms, plus a ten-minute idle window with zero findings. These are release tests, not a measured false-positive rate in production.

Known attacks, caught as behavior

EDAMAME has flagged the axios npm RAT, the tj-actions/changed-files compromise and the litellm PyPI takeover. Of the 12 publicly documented incidents in our September 2026 threat-coverage review, 11 exercise a technique it already detects.

Deterministic, no model needed

These detectors run on the device, with no data leaving the machine. A new indicator arrives as a signed threat model, with no app update. A model adds intent reading and explanations, as the table below shows.

The attack scenarios are public: test triggers on GitHub

What needs a model

Detection runs without a model. A model adds intent and explanations.

A model is optional. It reads each agent's transcripts to predict its intent for the divergence check, reduces noise and explains findings.

CapabilityWhat it doesNeeds a model
AI agent discoveryFinds the agents, MCP servers and harnesses on the deviceNo
Posture checksScores the device against the threat modelsNo
Traffic anomaly detectionFlags unusual network behavior per processNo
Attack-pattern detectionCredential harvest, token exfiltration and the other deterministic checksNo
Intent divergenceCompares what the agent declared with what the machine didPartly. A deterministic check runs without one; a model reads the transcripts to predict intent and refine the verdict
Noise reductionAdjudicates context on findingsYes. It can never clear a corroborated critical finding
Explanations and fixesThe AI Security Assistant explains findings and proposes remediationYes
Access decisionsEDAMAME Hub applies your policy to the reported postureNo

Model choices: EDAMAME's managed AI service in EDAMAME Portal (free tier included), your own Claude or OpenAI key, or a local model through Ollama.

Every agent on the machine

The agents you already run, observed as they are.

CursorClaude DesktopClaude CodeCodexOpenClawHermes
EDAMAME Security, Agents > Exposure: the agents on the machine, what each can reach and the findings that need attention
EDAMAME Security · Agents › Exposure (demo data)

Found from their transcripts

Agents are discovered from their transcript folders on disk: nothing to install for them to be seen.

What each one can reach

MCP servers, tools and secrets, and the agents that run unsandboxed, with passwordless root, or drive critical subprocesses.

Harness coverage

EDAMAME detects the sandbox or harness you already run and flags the agents running without one.

An observer the agent does not control

An agent cannot dismiss a finding or switch detection off: those actions stay with the operator, and a paused observer is itself reported.

Threat model

What the observer assumes.

Beside the agent, not inside it

EDAMAME reads processes, files and network from the operating system, outside the agent's own process, with no plugin inside the agent. On macOS this runs through the EDAMAME Helper, a system service.

The operator keeps the verdicts

Dismissing a finding, switching detection off and pausing the observer stay with the operator, not the agent.

A paused observer is reported

An agent on the machine whose observer is paused raises an unsecured-agent finding, and EDAMAME Hub shows it on the device's AI posture.

Coverage follows the host

EDAMAME observes the agents that run on its own machine. For an agent in a VM, a container or on a remote server, run EDAMAME Posture there.

Silence is visible

A device that stops reporting is escalated in EDAMAME Hub as signal lost.

Administrator rights are the line

The observer assumes the agent does not hold administrator or root rights. EDAMAME flags agents that can get root without a password, with the fix.

CI/CD runners and agent servers

The same proof on runners, build hosts and agent servers.

EDAMAME Posture is the CLI for CI/CD runners and any server running AI agents, cloud or self-hosted, such as an OpenClaw or Hermes host.

EDAMAME Hub, Endpoints > CI/CD: the runners reporting to the workspace with their security scores
EDAMAME Hub · Endpoints › CI/CD (demo workspace)

Posture before secrets

Patch, firewall, encryption and remote-access checks on the runner, before it touches a repository or a secret.

Egress allow lists

Each outbound connection is checked against the runner's allow list. A violation fails the build, and with cancel-on-violation it can stop the pipeline while it runs.

In your pipeline

Run it as a CLI, or in GitHub Actions with the EDAMAME Posture action.

Isolate from EDAMAME Hub

A critical finding takes the host out of your access lists.

EDAMAME Hub, Security > Policies: each policy with its compliant and non-compliant devices, including a policy for AI agents
EDAMAME Hub · Security › Policies (demo workspace)

Policies on posture and agents

A critical finding or a divergence verdict lowers the device score, so the host stops meeting your EDAMAME Hub policy.

Access follows the policy

When the host falls out of policy, EDAMAME Hub updates your access control (GitHub, GitLab, Microsoft Entra ID, Google, Tailscale, NetBird, Netskope or FortiGate), and that provider enforces it.

Fixed, back in

Once the issue is resolved, the device meets the policy again and gets its access back.

Mapped to the frameworks

Every AI check carries its OWASP Agentic, OWASP LLM, MITRE ATLAS and ISO/IEC 42001 references.

“Signed, allow-listed AI CLIs and runtimes become ‘living-off-the-AI’ tooling”

Microsoft adds that detection depends on “establishing what normal looks like for each agent [...] so that deviation becomes visible.” Source: Microsoft Digital Defense Report 2026, pp. 16–17.

Read the white paper

Read the security section of the white paper, deploy EDAMAME Security on workstations and EDAMAME Posture on runners and agent hosts to verify divergence across Cursor, Claude Desktop, Claude Code, Codex, OpenClaw and Hermes, detect attacks, and cut a compromised host’s access automatically.

EDAMAME Posture running headless on a CI runner: remediated threats and the security score after remediation

Read the white paper

Read the security section of the white paper, deploy EDAMAME Security on workstations and EDAMAME Posture on runners and agent hosts to verify divergence across Cursor, Claude Desktop, Claude Code, Codex, OpenClaw and Hermes, detect attacks, and cut a compromised host’s access automatically.

EDAMAME Posture running headless on a CI runner: remediated threats and the security score after remediation