An AI Agent Risk Scan Needs an Owner, Not Just a Score

A severity score does not fix an AI-agent risk. It does not prove the finding is exploitable, identify the affected business process, assign an owner, verify a mitigation, or show that production behavior is safe.

IBM’s new watsonx.governance capability is a constructive step: it scans connected repositories, maps potential agent risks to the OWASP Top 10 for Agentic Applications, links findings to code, recommends mitigations, and can alert teams when code changes.

For operating businesses, the practical lesson is bigger than the product. Agent-risk discovery must feed a closed workflow that connects code, permissions, business consequences, release decisions, runtime evidence, and recovery.

What IBM announced on November 19

Documented: IBM says watsonx.governance can identify potential AI-agent security risks before runtime using the OWASP Top 10 for Agentic Applications framework. Findings include a severity level, recommended mitigation, and references to specific lines of code that support the finding.

IBM also says the service can discover files containing agent code in a connected repository, scan the repository on a recurring basis, detect changes, and alert the user to run another scan so the Governance Console reflects current information.

Documented capability versus vendor claim: IBM documents risk identification, code references, severity, recommendations, and recurring change detection. Its statements about improving collaboration and confidence are expected benefits, not independently established business outcomes.

What the announcement does not establish

Unresolved: IBM’s public announcement does not publish precision or recall measurements, false-positive rates, benchmark coverage by agent framework, time-to-remediation results, or evidence that a recommended mitigation blocks the relevant behavior in production.

Static analysis can see code and configuration that are available to the scanner. It may not see effective cloud permissions, runtime secrets, dynamically loaded tools, external MCP servers, changed model behavior, retrieved instructions, poisoned memory, or downstream controls in ERP, commerce, manufacturing, and finance systems.

OWASP describes its Top 10 as a practical starting point. OWASP’s own exploit roundup also distinguishes verified facts from analyst mappings and repeatedly notes where full impact, CVEs, or mechanisms remain unknown. That is the right posture: use the taxonomy to organize investigation, not to declare certainty.

Independent reporting from ITPro has highlighted visibility and containment gaps using vendor-sponsored survey data. Those figures are useful context, but they remain reported survey claims rather than universal measurements.

A finding is a hypothesis until the business verifies it

A useful finding should answer five questions:

  1. Asset: which agent, version, repository, model, tools, credentials, and environment are involved?
  2. Path: what input, dependency, memory, tool, or permission could produce the risky behavior?
  3. Consequence: which customer, order, payment, price, file, machine, work instruction, or employee decision could be affected?
  4. Control: what independent boundary prevents or detects the action today?
  5. Evidence: which test and runtime records demonstrate that the boundary actually holds?

Without those answers, a dashboard can create motion without reducing risk. Connect security findings to AI governance and the responsible business system, not only to a developer backlog.

Build a closed finding-to-proof workflow

  1. Discover: inventory the agent and record its owner, purpose, model, tools, data, credentials, dependencies, and business process.
  2. Triage: confirm the code reference, reproduce the risky path where practical, and classify uncertainty.
  3. Prioritize: combine technical severity with business impact, reachability, data sensitivity, transaction authority, and recovery difficulty.
  4. Remediate: change permissions, validation, tool design, isolation, approval logic, monitoring, or workflow boundaries.
  5. Retest: run the original abuse case plus regression cases and expected legitimate work.
  6. Release: require evidence-based approval before the changed agent returns to production.
  7. Monitor: verify that production behavior and effective permissions remain inside the tested boundary.

Record every state change, owner, due date, decision, exception, evidence link, and retest result. This turns a finding into an auditable operating workflow.

Prioritize by consequence, not score alone

A medium technical finding can be urgent if the agent can release a purchase order, change a production recipe, expose a customer account, or send a binding commitment. A high score may be less urgent when the path is unreachable in production and protected by independent controls.

  • Commerce: catalog changes, negotiated prices, order edits, refunds, payments, tax, customer messages, and fulfillment instructions.
  • Industrial operations: recipes, setpoints, work instructions, quality dispositions, maintenance actions, and equipment access.
  • Distribution: inventory movements, vendor changes, purchase orders, allocations, shipments, and return authorizations.
  • Corporate systems: identity, payroll, finance, files, email, CRM, and administrative access.

Score business impact with the accountable operations owner. Use commerce controls and industrial-automation controls where the consequence actually occurs.

Reduce reachable authority before tuning the model

The fastest durable mitigation is often outside the model.

  • Give each agent a named identity and least-privilege access scoped to approved tools, records, environments, methods, amounts, and operating windows.
  • Separate read, propose, approve, execute, reverse, and administer permissions.
  • Validate stable record IDs, status, quantity, price, destination, revision, and authority limits deterministically.
  • Use allowlists for tools, endpoints, repositories, network destinations, and MCP servers.
  • Keep secrets out of prompts, memory, logs, repositories, and tool results.
  • Isolate parsers, code execution, browser activity, file handling, and external content from high-value systems.
  • Apply retry, chain-depth, token, cost, rate, and transaction limits.

Bridge Road’s agent cost and access governance provides the operating pattern. A safer prompt is useful; a permission the agent never receives is stronger.

Make remediation pass a release gate

Closing a ticket because code changed is not enough. The release record should connect the original finding to the exact agent version, changed files and policies, test fixtures, expected results, approver, deployment, rollback package, and monitoring plan.

  • Re-run the original exploit or abuse case.
  • Test prompt override, tool misuse, privilege escalation, memory poisoning, data exfiltration, recursive calls, approval bypass, and multi-agent chaining where applicable.
  • Confirm authorized work still succeeds; a control that blocks everything is not a workable production control.
  • Require a new review when tools, permissions, prompts, memory, retrieval, model provider, MCP dependencies, or workflow authority change materially.
  • Bind human approval to the exact action and parameters for high-impact or irreversible work.

OWASP’s agent-security guidance calls for structured testing before production and after material changes. Connect that gate to Bridge Road’s human approval controls.

Verify the control at runtime

Static findings identify possible design risk. Runtime monitoring shows which paths are actually exercised and whether the controls still hold.

  • Log agent, session, model, tool, policy, credential, approval, record, and transaction identifiers.
  • Monitor effective permissions, tool and MCP changes, unusual destinations, secret access, denial spikes, bypass attempts, repeated retries, long chains, and unexplained cost growth.
  • Verify consequential results against the system of record rather than the agent’s summary.
  • Protect evidence from the agent’s write and delete permissions.
  • Alert when repository, dependency, model, tool, policy, or credential changes invalidate the last assessment.
  • Measure time to acknowledge, contain, remediate, retest, and close each finding.

A scan result and a runtime event are different evidence. Preserve both so investigators can connect the suspected weakness to actual behavior and the final business outcome.

Assign ownership before the first finding arrives

Security owns the taxonomy and investigation method; development owns code changes; the agent owner owns behavior and deployment; operations owns the business queue; the system owner verifies destination records; and a business leader accepts residual risk.

Define a response for confirmed exploitation, failed controls, missing evidence, and uncertain outcomes:

  1. Pause or reduce the agent’s permissions.
  2. Preserve code, configuration, prompts, tool definitions, credentials, logs, approvals, and affected records.
  3. Check the destination system before retrying or reversing anything.
  4. Quarantine affected orders, files, work instructions, messages, or transactions.
  5. Restore the last known-safe version or manual process.
  6. Correct through an approved workflow, retest, and restart with reduced scope plus enhanced monitoring.

Connect this to the AI incident-response workflow and include the scanner itself in the vendor and model risk review.

A practical 30-day finding-to-proof exercise

  1. Select one agent that can reach a real business system but performs reversible, low-risk work.
  2. Inventory its code, tools, models, data, credentials, MCP dependencies, owner, and manual fallback.
  3. Map the design to the OWASP Top 10 for Agentic Applications and run a static scan.
  4. Choose one credible finding and document reachability, consequence, existing controls, uncertainty, and owner.
  5. Reduce permissions and add deterministic validation before changing model instructions.
  6. Create a repeatable abuse case, remediate, retest, and confirm legitimate work still succeeds.
  7. Deploy through a human release gate with monitoring, rollback, and an exception queue.
  8. Verify production evidence and close the finding only when the control—not merely the code change—is demonstrated.

Direct sources

Turn findings into verified controls

Bridge Road helps operating businesses connect agent inventory, permissions, findings, remediation, approval gates, runtime evidence, monitoring, and recovery around real workflows.

Schedule a 30-minute friction review