Skip to content
Blog

Field notes on agent security

Working notes on governing AI agents across frameworks and harnesses at the endpoint: shadow MCP, endpoint governance, and verifiable custody. No fluff.

Topics:agent-detection-and-responseagent-governanceagent-securityagentic-sdlcapprovalsauditautomation-templatesclaude-codecloud-runtimecodexcoding-agentscompliancecredentialscursoredrendpointendpoint-securityengineeringenterpriseevidencegatesgatewaygithub-copilotgovernanceinventorylatencylyramaturitymcpmeasurementobservabilityopentelemetrypolicypolicy-as-codeprompt-injectionregosdlcsecretssecurity-operationssemantic-policyshadow-mcpsiemsocspectrastatioumbra
  • 11 min read

    What the research says about prompt injection and tool misuse in AI agents, and how to contain it

    The measured failure rates for prompt injection and wrong tool calls, what MCP adds to the risk, and the three policy layers Alyria enforces locally at every tool call to contain what the model cannot.

    agent-securityprompt-injectionmcp
    Brian McManusRead →
  • 10 min read

    SDLC automation templates for AI agents: a catalog of what you can automate and govern

    A catalog of SDLC automation templates for AI agents: what each agent automates at the plan, design, build, test, review, release and operate stages, the guardrail that scopes it, and the evidence it leaves behind.

    agentic-sdlcautomation-templatesagent-governance
    Brian McManusRead →
  • 8 min read

    Agentic AI maturity for security operations teams

    Security operations maturity for AI agents is how far agents run the prepare, detect, investigate, respond, recover and improve loop, and whether runtime policy can stop a high-impact response action that the incident does not justify.

    maturitysecurity-operationssoc
    Brian McManusRead →
  • 8 min read

    The five signs your AI agent rollout has outgrown its controls

    Five signs that agent adoption has moved past the controls around it: no real MCP inventory, policy that lives in prompts, enforcement before a monitor period, evidence scattered across consoles, and approvals in the wrong place. Each one with the check to run.

    maturitygovernancecoding-agents
    Brian McManusRead →
  • 8 min read

    Laptop or cloud: where should your coding agents run?

    Run each coding agent where the work needs to happen, and put the same policy, secrets and audit in front of it either way. What a laptop gives you, what a governed cloud runtime gives you, and the questions that decide which run goes where.

    cloud-runtimestatiocoding-agents
    Brian McManusRead →
  • 7 min read

    Human approvals for AI agents without approval fatigue

    Require a person only where the action is irreversible or expensive, hold that action and nothing else, and record the approval next to the policy decision that asked for it.

    approvalsgovernancecoding-agents
    Brian McManusRead →
  • 9 min read

    Agentic AI maturity in the SDLC: from copilots to autonomous coding agents

    SDLC maturity for AI agents is the number of the six lifecycle phases an agent runs end to end, plus the controls that decide its actions in each one. Where copilots stop, what an autonomous coding agent needs, and the order to build it.

    maturitysdlccoding-agents
    Brian McManusRead →
  • 8 min read

    Gates for coding agents: where to stop an agent before push, PR and deploy

    Checks in the hook path have microseconds of budget. Checks at push, pull request and deploy can wait. A stage map of deterministic gates for coding agents, with the sequence policies that make a cheap check smarter than it looks.

    policygatescoding-agents
    Brian McManusRead →
  • 8 min read

    You already run EDR. Here is what it can't see about coding agents

    Endpoint detection and response watches processes, files and connections. A coding agent's tool calls happen inside one process, so the event that matters never reaches the sensor. What EDR covers, where it stops, and what to add for coding agent endpoint security.

    edrendpoint-securitycoding-agents
    Brian McManusRead →
  • 8 min read

    Evidence for AI-written code: what to record from prompt to merge

    Which agent, which model, which prompt, which tool call, which files, which decision, which approval. The eight fields that let a reviewer reconstruct an AI-written change without trusting a screenshot.

    auditevidencegovernance
    Brian McManusRead →
  • 8 min read

    Sending coding agent telemetry to your SIEM with OpenTelemetry

    What Claude Code, Codex, Cursor and Copilot each emit, the collector configuration that receives it, the exact attributes to delete before storage, and the detections that only work once the decisions and the tool calls land in the same index.

    observabilityopentelemetrysiem
    Brian McManusRead →
  • 8 min read

    Governing AI agents across the SDLC: from prompt to merged PR

    A coding agent can act in eight places between a prompt and a merged pull request. What governance means at each stage, which control can decide there, and where the record of the decision ends up.

    governancesdlccoding-agents
    Brian McManusRead →
  • 8 min read

    Short-lived credentials for AI agents: leasing instead of handing out keys

    Per-task, expiring credentials and the mechanics behind them: AWS session policies and credential_process, OIDC for CI, and where a lease has to come from when the agent runs on a laptop. Includes the bootstrap problem, which is the part teams miss.

    secretscredentialsagent-security
    Brian McManusRead →
  • 9 min read

    How fast can semantic policy be? Latency budgets for governing agent actions

    A governance layer cannot add a second to every tool call. The budgets each tier of policy can hold: deterministic checks in the hook path, sequence state, a small local model, and the boundaries where a slow check is free.

    semantic-policylatencypolicy-as-code
    Brian McManusRead →
  • 9 min read

    How to measure agentic AI maturity in your engineering org

    Measure agentic AI maturity on two axes: how many workflows an agent runs end to end, and how well those agent actions are controlled. A method you can run in a week using evidence you already have, plus the numbers worth collecting.

    maturityengineeringmeasurement
    Brian McManusRead →
  • 9 min read

    What an audit trail for AI coding agents has to prove

    Four questions every agent audit record must answer, what each coding agent logs natively, and what a hash chain does and does not establish.

    auditcomplianceagent-security
    Brian McManusRead →
  • 9 min read

    Keeping coding agents away from .env files, SSH keys and cloud credentials

    The deny rules, ignore files and sandboxes each coding agent offers for secrets, where they stop, and the arrangement that removes the credential instead of hiding it.

    secretsagent-securityendpoint
    Brian McManusRead →
  • 8 min read

    Sequence policies for AI agents: after reading a secret, block network egress

    A sequence policy decides on an action using what already happened in the session. The state model behind it, a Rego rule you can test, where each agent client lets you keep that state, and the failure modes that make these rules fire on the wrong thing.

    policypolicy-as-coderego
    Brian McManusRead →
  • 9 min read

    How to block a coding agent's tool call before it runs

    Pre-tool hooks in Claude Code, Cursor and Copilot, the sandbox and approval controls in Codex, and what each one can and cannot stop.

    agent-securitymcppolicy
    Brian McManusRead →
  • 8 min read

    GitHub Copilot MCP allowlists: what managed settings cover and what they miss

    How allowedMcpServers and deniedMcpServers match servers, how the deployment methods combine, which Copilot surfaces the allowlist skips, and the local activity GitHub's audit log does not record.

    github-copilotenterprisemcp
    Brian McManusRead →
  • 9 min read

    Securing Claude Code in the enterprise: managed settings, hooks and what they don't cover

    Managed settings, managed MCP configuration and hooks give an administrator real control over Claude Code on a laptop. Here is how each works, and the four things they leave to you.

    claude-codeenterprisepolicy
    Brian McManusRead →
  • 8 min read

    Codex requirements.toml for security teams: approvals, sandbox modes and MCP allowlists

    The admin-enforced layer in OpenAI Codex, key by key: where requirements.toml is read from, how sandbox and approval allowlists work, how MCP server identities are matched, and what the network proxy does not filter.

    codexenterprisemcp
    Brian McManusRead →
  • 8 min read

    Can an AI gateway control what coding agents do on a laptop?

    A gateway governs the traffic routed through it. Shell commands, file reads, and stdio MCP calls never take that path, so the question is which layer each control sits at.

    mcpgatewayendpoint
    Brian McManusRead →
  • 8 min read

    Cursor enterprise security settings: MCP allowlist, hooks, sandbox and the gaps

    How Cursor's MCP allowlist matches servers, where hooks and run modes fit, and the four things an administrator still owns after the settings are deployed.

    cursorenterprisemcp
    Brian McManusRead →
  • 8 min read

    How to find every MCP server on your developer laptops

    An MCP server inventory means reading each client's config on each machine, then keeping that read current. Here are the paths, the manual sweep, and where the sweep stops working.

    mcpshadow-mcpinventory
    Brian McManusRead →