Skip to content
VeriSwarm
About
DocsPricingAgent Skill
LoginRegister
  1. Home
  2. /Owasp
VeriSwarm
  • ✓ English
  • Español
  • Deutsch
  • Français
  • Italiano
  • Português
  • 日本語
  • 한국어
  • 简体中文

Product

  • Pricing
  • Docs
  • API
  • Agent Skill
  • OATS Spec

Trust

  • Trust Center
  • Security
  • Compliance
  • Status
  • Changelog

Company

  • About
  • Blog
  • Open Source
  • Investors
  • Press

Legal

  • Terms
  • Privacy
  • SLA
  • DPA
  • Accessibility

Security Attestation

OWASP Agentic AI Top 10

The OWASP Top 10 for Agentic Applications (2026) defines the most critical security risks for AI agent systems. Here’s how VeriSwarm maps to each one. For the top two risks in depth, see how VeriSwarm approaches prompt injection detection and MCP tool poisoning.

7/10
Risks Covered
3
Partial
0
On Roadmap
ASI01

Agent Goal Hijack

CoveredGuard + Gate

Untrusted ingested content (documents, web pages, email, tool responses) redirects the agent's objectives while it believes it still serves the user — agentic prompt injection at the planning layer.

  • Guard injection detection (DeBERTa ML + structural analysis)
  • Guard Proxy runtime interception
  • SOUL security scanning on templates
  • Cortex Workflows human_review gate for goal-changing actions
  • Trust-scoring anomaly detection
ASI02

Tool Misuse & Exploitation

CoveredGuard

Agents invoke legitimate tools harmfully — bad parameters, unsafe tool chaining, forwarding unvalidated output to a more powerful API — or are steered by poisoned tool metadata.

  • Guard tool permissions (per-tool allow/deny-list)
  • Guard Proxy runtime policy enforcement on every call
  • MCP scanner (poisoned tool metadata detection)
ASI03

Identity & Privilege Abuse

CoveredPassport + Gate

Excessive, shared, long-lived, or inherited credentials enable credential reuse, privilege escalation, and agent impersonation.

  • Passport Ed25519-signed agent manifests
  • ES256 1-hour portable credentials
  • Scoped, revocable delegations
  • Agent identity registry
  • Trust tier enforcement
  • Cedar policy engine
ASI04

Agentic Supply Chain Vulnerabilities

CoveredGuard + Marketplace

Compromise via frameworks, model providers, tool integrations, MCP servers, or registries — package poisoning, MCP CVEs — some only discovered at runtime.

  • SOUL security scanning (injection, bypass, content moderation)
  • MCP scanner (real MCP Top 10 checks)
  • Ed25519 template signing with signature-verified import
  • Template marketplace per-tenant isolation
ASI05

Unexpected Code Execution (RCE)

PartialGuard + Cortex

Natural language turns into running code or shell commands outside intended boundaries — the wall between instruction and execution collapses, enabling sandbox escape or RCE on host or downstream systems.

  • Guard tool permissions can deny code-exec/shell/SQL tools
  • MCP scanner command-injection detection
  • Cortex Workflows human_review gate before prod-modifying code runs

VeriSwarm restricts code-execution tool access and scans for command injection but does not sandbox agent runtime execution; deploy isolated/ephemeral execution runtimes for full ASI05 coverage.

ASI06

Memory & Context Poisoning

PartialCortex + Vault

Adversaries plant false or malicious information into session context, retrieval indexes, or persistent memory, shaping agent behavior across future sessions and users.

  • Cross-model verification (majority consensus across multiple LLMs)
  • Context governance (topic trending, context quality, knowledge-gap detection)
  • Guard input scanning
  • Vault provenance logging

VeriSwarm verifies critical decisions and logs provenance but does not manage the agent's memory store; validate memory writes and audit persistent memory in your agent runtime.

ASI07

Insecure Inter-Agent Communication

CoveredA2A + Passport

Agent-to-agent messaging lacks authentication, integrity, or origin verification, enabling spoofing, tampering, replay, or fake-peer registration — one compromised agent pivots across the network.

  • A2A transport Ed25519 message signing
  • Hop enforcement (deny/reaper)
  • Trust-ranked catalog (excludes killed agents)
  • Portable ES256 JWT credentials
  • Agent cards with x-veriswarm-trust extension
ASI08

Cascading Failures

CoveredCortex + Guard

One bad decision, hallucination, or compromise propagates through connected agents and workflows faster than humans can detect — one agent's output becomes another's trusted input.

  • Kill switch (immediate agent disable)
  • 3-state circuit breakers with SLO/error-budget enforcement
  • LLM fallback chains with provider health tracking
  • Hop enforcement
  • Cortex Workflows budget enforcement
ASI09

Human-Agent Trust Exploitation

PartialGate + Guard

The human is in the loop and does approve, but the agent manipulates them into it — fluent, confident summaries that hide the real underlying operation, or persuasion to extract credentials or data. Logs show valid human approval while what was displayed does not match what executed.

  • Article 50 disclosure (AI-status + principal identity)
  • Trust badges, attribution, and scoring
  • Guard content moderation
  • SOUL scanning for deceptive/impersonation personas

VeriSwarm provides AI-status and principal disclosure (Article 50 readiness), trust attribution, and deceptive-persona scanning; it does not control the agent's UI to enforce raw-operation-vs-summary display integrity.

ASI10

Rogue Agents

CoveredGuard + Fleet

The agent itself becomes an insider threat — loss of policy containment by an actor already inside the trust boundary, via design flaw, drift over time, or compromise. It quietly pursues unauthorized goals and may persist across sessions while surface activity looks normal.

  • Kill switch (named, suspends in minutes)
  • Trust scoring + reaper (kills low-trust/drifted agents)
  • Moderation and anomaly flags vs. a documented baseline
  • Session Sentry (multi-turn exfiltration detector)
  • A2A catalog excludes killed agents
  • Fleet lifecycle governance

Get your tenant attestation report

The /v1/compliance/owasp-attestation endpoint returns a per-tenant report with your specific coverage status, evidence counts, and upgrade recommendations.

Read the DocsStart Free