OWASP ASI07 Explained: Insecure Inter-Agent Communication
OWASP ASI07 (Insecure Inter-Agent Communication) explained: how agents trust each other too much, relay malicious instructions, and amplify prompt injection.
ReadAdversarial & AI Security.
AI security, MCP server reviews, and red-team write-ups across smart contracts, dApps, and Web2 infrastructure.
Showing 12 of 35
OWASP ASI01 Explained: AI Agent Goal Hijacking
OWASP ASI01 (Agent Goal Hijack) explained: how prompt injection redirects AI agent objectives. Direct, indirect, and tool-mediated patterns with mitigations.
Read
Agentic DeFi security: when AI agents control treasury, trading, and liquidations
AI agents now autonomously control DeFi treasuries, execute trades, and trigger liquidations. The cross-layer attack surface that contract audits cannot see.
Read
CVE-2025-49596: Anthropic MCP Inspector RCE Explained
CVE-2025-49596 (CVSS 9.4 Critical): unauthenticated RCE in Anthropic's MCP Inspector. How the proxy architecture failed, the patch (v0.14.1), and lessons for MCP dev tools.
Read
Long-Lived Agents: Delayed Execution Risk
Why long-lived AI agents fail across time, not just prompts. Practical audit checks for delayed execution, stale approvals, and memory-driven authority drift.
Read
OWASP ASI05 Explained: AI Agent RCE Patterns
OWASP ASI05 (Unexpected Code Execution) explained: how agent-generated code and tool composition produce RCE in agentic systems. Real CVEs and mitigations.
Read
Cursor IDE MCP CVEs: MCPoison & CurXecute Explained
CVE-2025-54136 (MCPoison) and CVE-2025-54135 (CurXecute): tool descriptor injection and workspace-file-write RCE in Cursor IDE's MCP layer. Mechanism, impact, fixes.
Read
The Web2 blind spot: Why audited smart contracts get hacked
Bybit, BadgerDAO, Curve — all audited, all drained off-chain. How DNS hijacks, CDN compromises, and signing-flow attacks bypass smart contract audits.
Read
AI Agent Outbound Authority: Audit Checks
Why email, messaging, and webhook tools need destination-level controls in AI agents. Practical audit checks for exfiltration and approval bypass.
Read
Indirect prompt injection: the Web3 agent attack chain
How indirect prompt injection drains Web3 agent wallets, poisons AI audits, and abuses MCP servers. Bankrbot case study and the auditor's 12-point checklist.
Read
Anthropic MCP SDK Vulnerability (April 2026): Full Analysis
Inside the April 2026 Anthropic MCP SDK design flaw: STDIO transport allows config-to-command-execution across Python, TypeScript, Java, Rust SDKs — by design.
Read
AI Agent Approval Bypass: Audit Checks
Why human approval fails in AI agents when the model still controls risky parameters. Audit checks for coding agents, long-lived agents, and Agentic DeFi.
Read