Skip to main content
The AI assessment assigns each risky call one primary category and up to two secondary categories. Each category belongs to one of six domains. The assessment also rates severity and evidence strength. Likely benign calls have no category or severity.

Categories

A category describes the call’s action and potential effects. Investigate the call to establish the agent’s intent and whether the call succeeded. A permission change on a test resource shares a category with one on your production cloud account. Read the explanation and arguments to understand the scope.

Severity

The assessment rates each call’s potential impact as Low, Medium, High, or Critical. It rates the requested action, so a blocked call can still be critical.

Evidence strength

The assessment rates evidence strength as weak, moderate, or strong. This measures how directly the command, arguments, and session context support its conclusion. Redacted values, truncated commands, and missing context weaken the evidence. High severity with weak evidence means the assessment could not confirm a potentially serious risk.

ATT&CK, ATLAS, and OWASP references

Each category maps to its closest entries in MITRE ATT&CK Enterprise v19.2, MITRE ATLAS v5.6.0, the OWASP Top 10 for Agentic Applications 2026, and the OWASP Top 10 for LLM Applications 2025. Kontext uses a fixed mapping for each category, checked on 22 September 2026. The AI assessment does not choose the mapping. Each entry identifies the closest technique if the call were an attack. Many flagged calls perform legitimate work with broad effects, such as a force push. These calls may only loosely match an attack technique. The Risk tab of the tool call drawer lists the entries for a finding’s primary category under Framework references, each linked to the framework. Slack risk alerts include the first ATT&CK, ATLAS, and OWASP agentic entry.
  1. ATLAS has no staging technique. Kontext maps its collection techniques to sensitive data access.
  2. These calls usually perform legitimate actions with broad effects, such as a force push or a cloud resource change. ATT&CK matches only hostile changes.
  3. ATLAS covers infrastructure an attacker owns. It has no technique for a victim exposing its own service.
Kontext uses these OWASP entries:

Cause references for Merlin flags

When Merlin flags a call that does not fit the user’s request, Kontext adds a Cause · Merlin flag row under Framework references. This row describes why the agent may have tried the call. The ATLAS references are AML.T0051.001 for indirect prompt injection and AML.T0051.002 for triggered prompt injection. The OWASP references are ASI01 and LLM01.