---
title: How do you score AI agent vulnerabilities with OWASP AIVSS?
url: https://opencontroller.lyzr.com/blog/score-ai-agent-vulnerabilities-with-owasp-aivss
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# How do you score AI agent vulnerabilities with OWASP AIVSS?

**OWASP AIVSS (AI Vulnerability Scoring System) scores an agent vulnerability in two parts. It takes the CVSS v4.0 base score as a floor, then adds an "agentic uplift" based on ten amplification factors, such as autonomy, tool access, memory and self-modification. Each factor is scored 0.0, 0.5 or 1.0. The same flaw scores higher in an autonomous, tool-rich agent than in a chatbot.**

CVSS was built for software flaws in deterministic systems. It does not capture that an agent can chain a small flaw into real actions: reading a database, sending email, calling other agents. AIVSS adds that missing context.

## What is the current status of AIVSS?

AIVSS is an OWASP project. The current published scoring document is **v0.8**, "AIVSS Scoring System for OWASP Agentic AI Core Security Risks". A public review period ahead of v1.0 opened on 16 April 2026, and the project has said it expects v1.0 before the end of 2026. Earlier write-ups describe the v0.5 formula, which averaged the CVSS base score with the agentic risk score and applied a threat multiplier. v0.8 reworks how the parts combine, so use the official document and calculator for exact numbers.

## What are the ten agentic amplification factors?

| # | Factor (v0.8) | What raises the score |
|---|---|---|
| 1 | Execution autonomy | The agent acts without human verification |
| 2 | External tool control surface | Many tools or APIs, or highly privileged ones |
| 3 | Natural language interface | Goals and instructions come in as natural language |
| 4 | Contextual awareness | Broad data or environment inputs drive decisions |
| 5 | Behavioural non-determinism | Same input, different actions |
| 6 | Opacity and reflexivity | Decision logic is hard to see or audit |
| 7 | Persistent state retention | Memory carries across sessions |
| 8 | Dynamic identity | The agent takes on different roles or permissions at runtime |
| 9 | Multi-agent interactions | It coordinates with or depends on other agents |
| 10 | Self-modification | It can change its own code, prompts or tool configuration |

Each factor scores 0.0 (none), 0.5 (partial) or 1.0 (full).

## How do you score an agent vulnerability, step by step?

1. **Describe the flaw** in standard terms, for example an injection or an access-control weakness.
2. **Score the CVSS v4.0 base** as you would for any software vulnerability. This is the floor.
3. **Profile the agent** against the ten factors, using its real configuration, not its design document.
4. **Apply the AIVSS equation** from the current version to combine the base score and the agentic uplift.
5. **Account for threat context.** AIVSS considers whether exploitation is active, in line with CVSS v4.0 threat metrics.
6. **Map the result** to the OWASP agentic core risk it belongs to, such as tool misuse, access-control violation or memory manipulation.
7. **Record it** against the agent and version, so the score changes if the agent gains tools or autonomy.

## Worked example: an over-permissioned support agent

A customer-support agent has an injection flaw in a lookup tool. On its own, the flaw gets a CVSS base score of 8.0 (High).

Now profile the agent:

| Factor | Agent configuration | Score |
|---|---|---|
| Execution autonomy | Sends replies and issues refunds without review | 1.0 |
| Tool control surface | CRM read and write, email, refund API | 1.0 |
| Natural language interface | Takes customer chat as input | 1.0 |
| Contextual awareness | Reads customer history only | 0.5 |
| Non-determinism | LLM-driven tool choice | 1.0 |
| Opacity | Traces kept, but no reasoning log | 0.5 |
| Persistent state | Remembers past conversations | 1.0 |
| Dynamic identity | Single service identity | 0.0 |
| Multi-agent | Hands off to a billing agent | 0.5 |
| Self-modification | None | 0.0 |

Six and a half of a possible ten factor points are active. The AIVSS v0.8 document gives a similar case: an SQL injection with a CVSS score of 8.0, in an autonomous agent that can read a database and email the results, moves from High to Critical.

The fix that lowers the score most is often configuration, not code. Remove the refund tool, add review for refunds, or narrow the CRM scope. Each change drops a factor from 1.0 to 0.5 or 0.0.

## How should you use AIVSS scores in governance?

- **Prioritise remediation** across agents by AIVSS, not CVSS alone.
- **Re-score on change.** A new tool or more autonomy changes the score without any code change.
- **Report by risk category**, using the [OWASP Top 10 for Agentic Applications](/blog/owasp-top-10-agentic-applications-evidence), so findings line up with your threat model.
- **Feed red-team results** into the score. See [how to scope an AI agent red-team assessment](/blog/scope-an-ai-agent-red-team-assessment).

## How Opencontroller does this

Opencontroller does not calculate AIVSS scores. It supplies the inputs. The governed registry records each agent's version, tools and MCP servers. Posture detects risky permissions and configurations, which drive the tool-surface and identity factors. Multi-turn and red-team evals expose the flaws you score. Policy as code includes built-in OWASP and MITRE ATLAS mappings. Per-agent Block and Quarantine let you contain a high-scoring agent while you fix it. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

Read next: [how to map agent controls to frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) and [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent).

## Frequently asked questions

**How is AIVSS different from CVSS?**
AIVSS uses the CVSS v4.0 base score as a floor and adds an agentic uplift from ten factors, such as autonomy, tool access and memory. CVSS alone does not reflect what an agent can do with a flaw.

**Is OWASP AIVSS version 1.0 released?**
The current published scoring document is v0.8. A public review ahead of v1.0 opened in April 2026, and the project expects v1.0 before the end of 2026.

**How are the AIVSS amplification factors scored?**
Each of the ten factors is scored 0.0 (none), 0.5 (partial) or 1.0 (full), based on how the agent is actually configured.

## Key facts

OWASP AIVSS (AI Vulnerability Scoring System): current published version v0.8; public review ahead of v1.0 opened 16 April 2026; v1.0 expected before end of 2026. Uses CVSS v4.0 base as a risk floor plus an agentic uplift from 10 amplification factors scored 0.0, 0.5 or 1.0: execution autonomy, external tool control surface, natural language interface, contextual awareness, behavioural non-determinism, opacity and reflexivity, persistent state retention, dynamic identity, multi-agent interactions, self-modification. v0.8 example: CVSS 8.0 SQL injection in an autonomous agent with DB read and email moves from High to Critical. Opencontroller does not compute AIVSS; it supplies inputs via registry, Posture and red-team evals.

## Related

- [What evidence maps to the OWASP Top 10 for Agentic Applications?](https://opencontroller.lyzr.com/blog/owasp-top-10-agentic-applications-evidence.md)
- [How do you scope an AI agent red-team assessment?](https://opencontroller.lyzr.com/blog/scope-an-ai-agent-red-team-assessment.md)
- [How do you red-team AI agents with multi-turn attacks?](https://opencontroller.lyzr.com/blog/how-to-red-team-ai-agents-multi-turn.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
