The Harness Compiled. That Was the Easy Part: LLM Fuzzing
LLM agents can generate fuzz harnesses at scale, but compilation is only the first gate. Build an evidence pipeline for coverage, crashes, and triage.
-
-
The Sandbox Was Only a Dataclass: PraisonAI’s Policy Gap
PraisonAI exposed a dangerous gap between declared sandbox policy and enforced runtime controls. We trace the agent-to-shell chain, reproduce two mechanics safely, and map the fixes and detections that matter.
-
The Button Was the Payload: A2UI CVE-2026-10032
CVE-2026-10032 shows how an agent-generated button can turn declarative UI data into browser behavior. Trace the sink, safe lab, patch, and detections.
-
The Browser Changed the Address: DBHub DNS Rebinding to SQL
CVE-2026-61742 turned a browser into a network pivot against DBHub MCP. We reproduce the DNS rebinding flaw, compare 0.22.4 with 0.22.5, and map practical detection and defense.
-
The Model Returned Python: CVE-2026-61539 and Tool-Parser RCE
CVE-2026-61539 turned attacker-influenced Llama tool output into Python RCE inside Xinference. This analysis traces the parser boundary, validates the primitive in a safe local lab, and maps durable detection and hardening controls.
-
MCP Security: How to Secure Agents, Tools, and Servers
MCP security in practice: the main Model Context Protocol risks, from indirect prompt injection to excessive privileges, and a practical hardening framework for agents, clients, and servers.
-
Enterprise AI Attack Chains: Prompt Injection, Tool Abuse, Detection
Technical analysis of enterprise AI attack chains: indirect prompt injection, memory poisoning, tool abuse, MCP exposure, and detection.
-
AI Threat Modeling: Trust Boundaries for RAG, Agents, and MCP
Advanced guide to AI threat modeling for RAG, agents, and MCP, with trust boundaries, context provenance, and identity separation.
-
Attacking Multi-Agent Systems: Rogue Agents and A2A
A practical AI Red Team guide to A2A reconnaissance, canary agent registration, agent-card spoofing, and delegation auditing in an authorized lab.
-
Agentjacking: When Your AI Assistant Becomes a Weapon
Agentjacking exploits the blind trust AI assistants place in MCP data. One fake Sentry event and your agent executes malicious code.
-
MCP Security: Tools and Permissions in Practice
MCP Security in practice: a safe lab with fictional documents, a mocked local agent, controlled tool exploitation, permissions, schemas, and auditable logs.
-
Attacks on Embeddings and Vector Databases
Embedding attacks in practice: inversion, membership inference, semantic probing, evidence artifacts, and defenses for vector databases.
-
AI Recon: LLMs, RAGs, and Agents
Authorized reconnaissance for enterprise AI: map LLMs, RAG pipelines, agents, tools, IAM, APIs, cloud exposure, and evidence without drifting out of scope.
-
Autonomous Pentest Agent: AI, Red Team, and Attack Chains
Autonomous Pentest Agents can accelerate Red Team operations, but they also introduce risk around prompt injection, tool misuse, scope control, evidence quality, and detection.
-
AI as C2: 5 Risks and Defenses for Blue Team
AI as C2 is a real threat in enterprise environments. This guide covers 5 critical risks, detection indicators, and practical defenses for Blue Team.
