Tools

Auth and injection issues in A2A Agent Cards

Prompt injection in AI-agent GitHub Actions

Shadow-AI inventory drift in a repository

Prompt-injection resistance with canary payloads

Oversized tool results and context blowups

SSRF and destination risk in fetched URLs

Missing human approval on consequential tool calls

Runaway loops and retry storms

Risky install manifests for MCP servers

Weak OAuth settings in MCP server metadata

Least-privilege issues in MCP client configs

Taint from MCP arguments into dangerous sinks

Poisoning in persistent agent memory

Floating model aliases and mid-run switches

Dangerous LLM output before it is rendered

Sensitive file paths an agent touched

Code-execution risk in pickle model files

PII in agent transcripts

Secrets and authorization logic inside system prompts

System-prompt leakage in transcripts

Indirect prompt injection in RAG documents

Indirect prompt injection in tool results

Backdoors in AGENTS.md and rule files

Behavioral drift between agent runs

Token and cost overruns in agent runs

Unconstrained high-privilege tool parameters

Over-privileged GitHub tokens used by agents

Secrets flowing into later tool arguments

Dangerous shell commands an agent ran

Hallucinated and typo-squatted package names

Prompt-template injection in system prompts

Policy violations in agent tool calls

MCP tool drift and rug-pulls

Hidden instructions and rug-pulls in MCP manifests

Secrets leaked into agent transcripts

The lethal trifecta across an agent's tools

Poisoned or backdoored fine-tuning data

Security smells in AI-written code

Over-permissive coding-agent settings