FRAMEWORK March 22, 2026 9 min read

DASF v3.0 Is Here: 35 New Agentic AI Security Risks — Which Ones Are Your Agent Skills Already Violating?

Databricks released DASF v3.0 on March 20, 2026 — adding 35 new agentic AI security risks and 6 new mitigation controls to the Databricks AI Security Framework.

If OWASP's Agentic Top 10 is the checklist security teams reference, DASF v3.0 is the enterprise implementation guide data and ML engineering teams will use to operationalize it. The two frameworks complement each other — and both map directly to what SkillShield scans for.

What DASF v3.0 Covers

The 35 new risks are organized around three sub-components that map to how AI agents actually work:

13A: Agent Reasoning Risks

These target the agent's decision-making loop:

13B: MCP Server Risks (the tool interface)

These target the tools agents connect to:

13C: MCP Client Risks (the connection layer)

These target how agents connect to tools:

Inter-Agent Risks

The Lethal Trifecta

DASF v3.0 highlights a critical concept from Meta's "Agents Rule of Two" and Simon Willison's research: the Lethal Trifecta. The risk profile spikes when three conditions are present simultaneously:

  1. Access to sensitive systems or private data
  2. Processing untrustworthy inputs (external prompts, web content, emails)
  3. Ability to change state or communicate externally (execute code, send emails, modify databases)

When all three are present, an indirect prompt injection embedded in untrusted data can hijack the agent's full capability set. Remove any single leg — by scoping permissions, adding a human checkpoint, or validating intent before tool selection — and the attack chain breaks.

This framing is directly relevant to skill scanning: SkillShield identifies skills that create condition #1 (excessive permissions) and condition #2 (untrusted inputs through poisoned tool descriptions), helping you break the trifecta before it forms.

The 6 New Controls

ControlDASF IDWhat It Does
Least privilege for toolsDASF 5, 57, 64Scope agent permissions to immediate task
Human-in-the-loop oversightDASF 66Require human verification for high-stakes actions
Sandboxing and isolationDASF 34, 62Agent code runs in ephemeral, isolated environments
AI Gateway and GuardrailsDASF 54Monitoring, safety filtering, PII detection
Observability of thoughtDASF 65Capture why the agent acted — planning steps, reasoning
Supply chain scanningValidate tools and MCP servers before connecting

Where SkillShield Maps to DASF v3.0

Direct Coverage

DASF RiskRisk IDSkillShield Detection
Tool Poisoning13.18Prompt injection detection in tool descriptions and SKILL.md
Prompt Injection in Tool Descriptions13.16Direct scanning target — hidden instructions in metadata
Malicious Server13.26Known malicious database (533 blocked), typosquat detection
Client-Side Code Execution13.32Dangerous code patterns: eval, exec, shell invocation
Data Leakage via Tool13.30Hard-coded secrets detection, excessive permission flagging

Partial Coverage (Pre-Installation Layer)

DASF RiskRisk IDSkillShield Contribution
Memory Poisoning13.1Catches poisoned tool descriptions that inject false context
Intent Breaking13.6Detects cross-tool manipulation in skill definitions
Rogue Agents13.13Blocks malicious skills that could compromise agents

DASF v3.0 vs OWASP Agentic Top 10

DimensionOWASP Agentic Top 10DASF v3.0
AudienceSecurity teams, developersData/ML engineering, enterprise security
Format10 risk categories (ASI01-ASI10)35 specific risks + 6 controls
GranularityCategory-level guidanceRisk-ID-level with control mappings
MCP coverageASI04 names MCP supply chainDedicated 13B (server) and 13C (client) sections
Industry mappingStandaloneMaps to MITRE ATLAS, OWASP, NIST, CSA

Breaking the Lethal Trifecta with Pre-Install Scanning

The practical takeaway from DASF v3.0: scan your tools before they get access to your systems.

Every MCP server and agent skill you connect creates the potential for the Lethal Trifecta. SkillShield breaks leg #1 (excessive access) by flagging over-permissioned tools and breaks leg #2 (untrusted inputs) by catching poisoned tool descriptions — before the skill is installed.

npm install -g skillshield
skillshield scan ./SKILL.md

Scan Against DASF v3.0

Check your agent skills and MCP servers against the Databricks AI Security Framework. Free, instant results.

Start Scanning