Agentic AI security
AI agents act on your behalf. Make sure they can’t be turned against you.
Agents now read your email, call your APIs, write code, and move data with real credentials. That makes every document, web page, and tool they touch a potential way in. We help you adopt AI with the same discipline you apply to any other privileged system.
Questions we help you answer
- Which AI agents and tools are running today — sanctioned or not?
- What data and systems can each agent reach?
- Who approved the plugins, connectors, and MCP servers they use?
- What happens if an agent reads a malicious document or web page?
- Can you show customers and auditors that it’s under control?
The risks we test for
Mapped to the OWASP Top 10 for Agentic Applications (2026).
Goal hijack & prompt injection
Hidden instructions in content redirect what an agent is trying to do. ASI01
Tool misuse & code execution
Legitimate tools used in unsafe ways, or generated code run without guardrails. ASI02, ASI05
Identity & privilege abuse
Over-broad credentials reused, escalated, or passed between agents. ASI03
Agentic supply chain
Compromised plugins, prompt templates, models, or external MCP servers. ASI04
Memory & context poisoning
Tampered memory or retrieved context that skews future decisions. ASI06
Multi-agent & trust failures
Unauthenticated agent-to-agent messages, cascading errors, rogue agents, and over-trusting humans. ASI07–ASI10
What we do
A security program for your AI — not a one-time scan.
AI inventory & risk assessment
- Discover sanctioned and shadow AI use
- Classify agents by data access and autonomy
- AI risks added to your risk register
Architecture & threat modeling
- Map agent data flows, tools, and trust boundaries
- Threat model each high-impact workflow
- Prioritized design fixes
Guardrails & least privilege
- Scoped, short-lived credentials per agent
- Human approval for high-impact actions
- Sandboxing, egress controls, and output filtering
Red teaming & testing
- Prompt-injection and jailbreak testing
- Tool-abuse and data-exfiltration scenarios
- Retesting after fixes
AI governance
- AI policy and acceptable-use standards
- Model, vendor, and connector review process
- Management system aligned to ISO/IEC 42001
Monitoring & response
- Agent activity logging and alerting
- Kill switches and credential revocation
- AI-specific incident response playbooks
AIUC-1 readiness & certification
AIUC-1 — the agent standard your enterprise buyers are starting to ask for.
AIUC-1 is an independent certification standard for AI agents, published by the Artificial Intelligence Underwriting Company with a consortium that includes Stanford, MIT, Carnegie Mellon, Cisco, and Google Cloud. It covers six risk pillars — data and privacy, security, safety, reliability, accountability, and society — and is refreshed every quarter. We get you ready for it and stay with you through the audit.
Gap assessment
We measure your agents against the current AIUC-1 release and hand you a prioritized gap list, effort estimate, and realistic timeline.
Scope & Statement of Applicability
Developer, deployer, or both; which agents are in scope; which requirements apply; and every exclusion documented the way your auditor will expect.
Control implementation
The safeguards themselves — tool-call validation, MCP and agent-to-agent protocol security, agent identity separated from access, logging, and human approval gates.
Evidence packages
Technical, legal, and operational evidence in the verifiable form auditors now expect, organized by requirement rather than screenshotted after the fact.
Independent evaluation prep
Readiness for third-party testing against the AIUC-1 risk taxonomy — typically one to three thousand simulations — and remediation of what it surfaces.
Audit & annual upkeep
We coordinate with your accredited auditor, then keep the certificate alive: evaluations at least quarterly and a re-audit each year.
What certification involves
Four phases, typically four to ten weeks.
Starting from zero takes longer; an existing ISO 42001, SOC 2, or HITRUST program shortens it considerably, because much of the governance evidence already exists.
| Phase | Typical duration | What happens |
|---|---|---|
| Scoping & kick-off | 1–2 weeks | Your role, the in-scope agents, and a Statement of Applicability signed off by the auditor |
| Evidence collection | 3–5 weeks | Controls implemented and documented across technical, legal, and operational areas |
| Independent evaluations | 3–5 weeks | Third-party testing of the agents against the AIUC-1 risk taxonomy |
| Audit report & certificate | 1–3 weeks | The auditor reports; AIUC issues the certificate |
| Keeping it | Ongoing | Certificate valid 12 months, evaluations at least quarterly, controls re-audited annually |
AIUC issues the certificate and an accredited AIUC-1 auditor performs the audit. We prepare you and work alongside them — and where we act as an assessor, we keep that work separate from advisory, as independence rules require.
Where it fits
AIUC-1 alongside the standards you already answer to.
These are complements, not competitors. ISO/IEC 42001 certifies the management system around your AI. AIUC-1 validates the safeguards on specific agents through independent evaluation, and publishes crosswalks that map its requirements onto the NIST AI RMF and the EU AI Act. Running them together means one control set and one evidence base.
- ISO/IEC 42001 — governance and management system
- AIUC-1 — tested safeguards on named agents, re-evaluated quarterly
- NIST AI RMF and EU AI Act — obligations the crosswalks map onto
- SOC 2 and HITRUST — the enterprise security proof your buyers already request
Keeping pace with quarterly releases
AIUC-1 changes every quarter, and certified organizations have to keep up. Recent releases added controls for open-weight models, consolidated third-party testing into one requirement per pillar, tightened MCP and agent-to-agent protocol security, and separated agent identity from agent access.
We track each release and tell you what it means for your scope — before your auditor does.
Frameworks & regulations
Aligned to the standards your customers and regulators reference.
| Framework | What it covers | How we use it |
|---|---|---|
| AIUC-1 | Certification standard for AI agents across six risk pillars, refreshed quarterly | Readiness, evidence, evaluation prep, and audit support |
| OWASP Top 10 for Agentic Applications (2026) | The most critical security risks for autonomous AI agents | Threat modeling and red-team test plans |
| OWASP Top 10 for LLM Applications | Prompt injection, data leakage, and other LLM risks | Secure design reviews for LLM features |
| NIST AI RMF (AI 100-1) & Generative AI Profile | Govern, Map, Measure, Manage functions for AI risk | AI risk assessment and governance structure |
| ISO/IEC 42001:2023 | Certifiable AI management system | Policies, roles, and controls ready for certification |
| MITRE ATLAS | Adversary tactics and techniques against AI systems | Detection engineering and attack scenarios |
| EU AI Act | Transparency duties from August 2, 2026; high-risk obligations from December 2, 2027 or August 2, 2028 | Applicability analysis and readiness roadmap |
| HITRUST AI Security Assessment | AI-specific security controls for certification | Readiness for healthcare and vendor assurance |
| CCPA ADMT regulations | Notice, opt-out, and access rights for automated decision-making from January 1, 2027 | Privacy and AI governance working together |
Rolling out AI agents this year?
Let’s review what they can reach and how they could be misused — before an attacker does.