OWASP AIVSS: Building an Evidence-Backed Agentic Vulnerability Score
OWASP AIVSS extends vulnerability assessment with AI-specific, temporal, environmental, and modified metrics. A useful assessment preserves the evidence behind every selected value, names the deployment context, and separates routed observations from model, training, infrastructure, and governance evidence. This guide turns the score into a reviewable security record.

OWASP's AI Vulnerability Scoring System starts with familiar vulnerability characteristics, then adds AI behavior and deployment context alongside time and mitigation. It also covers impact. The official repository recommends its V4 calculator for new assessments and breaks the AI portion into 39 questions across nine metrics. That creates a richer record than a severity label, but only when every selected value carries evidence. I would reject a score delivered as a bare decimal. A number without the test result and system boundary is impossible to review. The same applies when the assessor's justification is missing. Such a score is cheap to manipulate.
TL;DR
- OWASP AIVSS is an incubator framework for scoring AI vulnerabilities, with V4 identified by the project as its recommended calculator for new assessments.
- A useful assessment records every metric selection and the evidence behind it. It also records deployment context and prioritized remediation, alongside the final score and risk category.
- Routed HTTP evidence can support caller and privilege observations, interaction and injection observations, disclosure and output observations, plus audit observations. Training, infrastructure, governance, and human-oversight claims need evidence from their actual owners.
- Treat the score as a living assessment record and decision input, rather than a substitute for security judgment.
AIVSS is a living OWASP incubator framework
The OWASP project page classifies AIVSS as an incubator framework. Its official specification repository begins with base vulnerability characteristics. It then adds AI-specific assessment and temporal factors. Impact measures and environmental requirements follow, along with modified values and mitigation. The same repository calls the framework a living document and says that it will be revised as research and threat intelligence change.
That status belongs at the top of an assessment. Record the repository commit or retrieval date beside the calculator version because a later revision may alter weighting as well as labels and values. The project currently identifies V4 as the recommended calculator for new work, while V1 and V2 are deprecated and V3 lacks V4's industry profiles and sub-category scoring, according to the implementation notes. Calling AIVSS a finalized standard or a replacement for every existing prioritization method would run ahead of the project's own claims.
The assessment record matters more than the decimal
The repository's reporting section requires the final score and risk category, detailed metric and sub-category scores, justifications tied to evidence, and vulnerability impact. It also requires prioritized mitigation recommendations and supporting material. Its implementation checklist also calls for reassessment after mitigations are applied, as documented in the AIVSS README. Those fields turn a calculator result into a reviewable decision record.
I would keep one record per vulnerability and deployment context. Put the affected model or agent and endpoint above the metric selections. Add the model version, tools, data classes, and environment. Then record the assessor and date, followed by the evidence references. Preserve each selected value with a short explanation. On a screen, the useful artifact looks like a stack of labeled rows. Each row shows the metric and chosen level, followed by the evidence link and owner. Any unresolved gap appears beside them. The decimal sits near the bottom because the reasoning produces it.
The minimum record should contain:
- Scope: the affected AI system and deployment, plus the version, vulnerability, and security boundary.
- Base assessment: attack vector and complexity, followed by privileges, user interaction, and scope, using the project's defined values.
- AI assessment: all nine current V4 metric groups, with their sub-category answers.
- Context: temporal and impact selections, plus the environmental, modified, and mitigation selections that apply to this deployment.
- Justification: evidence for every selected value, plus the owner of any missing evidence.
- Decision: final score and risk category, followed by remediation order, acceptance owner, and reassessment date.
These fields follow the components and reporting expectations in the official AIVSS specification.
Nine AI dimensions force separate evidence questions
V4 asks 39 sub-category questions across nine AI metrics. The official implementation details covers model resilience and data sensitivity. It also covers ethical impact, decision criticality, adaptability, and agentic autonomy. LLM-specific risks are included. Governance and visibility appear alongside contextual sensitivity. Averaging sub-categories is intended to stop one moderate top-level choice from hiding a severe component.
Keep those groups separate during evidence collection. A prompt-injection test can support an LLM-risk selection. It says little about training-data provenance. An authorization test can show that an agent exceeded its delegated role. It cannot prove that an incident-response process works. A fairness review and model card answer different questions from a threat model or red-team result. Policy records and routed request logs provide other evidence. Compressing them into a single screenshot creates false confidence.
The OWASP agentic controls mapping is useful for assigning control owners after a weakness is identified. The OWASP agentic audit-evidence guide covers the evidence packet behind those controls. AIVSS adds a scoring and prioritization layer; it still depends on those underlying artifacts.
Environmental scoring belongs to the deployed system
AIVSS environmental metrics include confidentiality and integrity requirements, plus availability and societal-impact requirements. Modified base metrics allow the deployment environment to change attack vector and complexity, or privileges, interaction, and scope, as specified in the official README. V4 also supplies seven industry profiles with different component weights and mitigation ranges, according to the implementation notes.
Choose those values against the system being assessed. A customer-support agent processing public product documentation has a different confidentiality requirement from a clinical agent receiving PHI. An assistant that drafts text for review has a different decision-criticality profile from an agent authorized to submit a transaction. Write the exact data class and action beside the selection, along with the approval path. "High impact" is a label. "Can send a refund without human approval" is an assessable condition.
Environment also changes over time. A new model version or added tool can invalidate earlier answers. The same applies to a broader role or changed data source. Direct provider fallback can also invalidate them. The repository's reassessment step is therefore operationally useful: reopen the record after material changes and after mitigation, preserving the earlier result rather than overwriting it.
Routed HTTP evidence has a strict stop line
For an authenticated user or agent calling an LLM over a deliberately routed HTTP endpoint, request and response records can support several AIVSS inputs. They can show the caller identity supplied by the application and the role or privilege context. They can also show user interaction, the selected endpoint, tested prompt-injection behavior, and sensitive-data exposure. The timestamp completes a routed record that already contains response handling and the policy decision. The LLM security testing guide explains how repeatable fixtures create stronger evidence than an unrecorded manual prompt.
That evidence has limits. An HTTP route cannot establish training-data integrity or model-weight provenance. It also cannot establish host security, cloud configuration outside the path, lifecycle governance, fairness across a population, or the effectiveness of human oversight. Local tool execution and STDIO traffic sit outside the route as well. Assign each unsupported AIVSS input to the model or MLOps owner, the cloud or application owner, the governance owner, or the human-process owner that can produce the evidence.
A score should expose those gaps. Use "evidence unavailable" in the justification and avoid turning a default value into an observed fact. The AIVSS repository itself warns that the framework is no guarantee of security and should not be the sole basis for security decisions, in its calculator disclaimer.
Remediation order needs the metric-level record
AIVSS says assessment findings should inform prioritized remediation and then be reassessed, according to its implementation checklist. The component record makes that practical. Two vulnerabilities can land in the same risk category while demanding different work. One may be dominated by broad agent authority and weak interaction controls. Another may come from sensitive data and prompt injection, with absent response filtering.
Write remediation beside the contributing metric and name the accepting owner. Reduce tool authority and add a human approval step. Restrict a route or remove sensitive fields. Add a regression fixture or close a direct-endpoint bypass. Then rerun the same fixture and attach the new evidence to the same assessment lineage. The score change becomes auditable because the underlying condition changed.
DeepInspect
DeepInspect contributes evidence for the authenticated HTTP route between a user or agent and an LLM endpoint. The application supplies identity and role context. DeepInspect evaluates the routed prompt and response under the active policy, then records the decision and classifications. It also records the endpoint and timestamp outside the calling application's write path, as described in the AI API security vendor guide.
That record can support selected AIVSS inputs and their metric justifications. Model training and cloud infrastructure retain their own evidence owners. The same applies to local execution, tool-side effects, governance, fairness, and human oversight. An honest AIVSS packet shows both sets clearly.
Book a technical deep dive at deepinspect.ai.
Frequently asked questions
- Is AIVSS a finalized OWASP standard?
The OWASP project page labels AIVSS an incubator framework. The repository also describes it as a living document that will be revised. Use it with a recorded version or commit, and avoid presenting its current calculator as a frozen industry standard.
- Which AIVSS calculator should a new assessment use?
The project's implementation details recommend V4 for new assessments. They describe V4 as adding seven industry profiles, plus 39 sub-category questions across nine AI metrics. Preserve that version in the assessment record so later reviews use the same reference.
- Can HTTP gateway logs fill every AIVSS metric?
No. Routed records can support observations about the authenticated call and prompt. They can also support observations about the response and policy decisions, with their timestamps. Training and weights need separate evidence, as do infrastructure and organizational governance. Societal impact and human oversight also need separate evidence. The score should identify those owners instead of inferring their condition from traffic.
- Should teams use AIVSS as the sole prioritization method?
The official AIVSS disclaimer says the framework is no guarantee of security and should not be the sole basis for security decisions. Pair the score with exploit evidence and business context, plus affected assets, remediation cost, and the organization's risk process.