AI Security
LLM05 improper output handling: a worked example and fixes
OWASP LLM05 means trusting model output as if it were safe. See a worked example of a chat feature that runs model SQL and how to fix it with controls.
Oct 18, 20264 min read
Resources
Supply Chain Security, in plain English.
Deep dives, practical guides, and incident analyses from engineers who build Safeguard. No fluff, no vendor FUD — just what you need to ship secure software.
Filter
Articles
RSS feedAI Security
LLM04 data and model poisoning: signals to log and alert on
Poisoning leaves traces in data provenance, training metrics and behavior. A defensive guide to the events worth logging and the alerts worth setting.
Oct 18, 20264 min read
AI Security
How does OWASP LLM04 data and model poisoning map to compliance?
OWASP LLM04 maps itself to no regulation, so here is a working map to EU AI Act Articles 10 and 15, NIST AI RMF and MITRE ATLAS, plus evidence to keep.
Oct 18, 20265 min read
AI Security
How to test an LLM app for sensitive information disclosure
OWASP LLM02 covers PII, secrets and business data leaking through model output. A defensive test plan with canary data, role checks and log reviews.
Oct 18, 20264 min read
AI Security
LLM02 sensitive information disclosure: which controls work?
OWASP LLM02 says sanitize data, limit access and set clear policies. A system prompt rule alone fails. Here is what to put in place first and why.
Oct 18, 20264 min read
AI Security
LLM01 prompt injection: how it works, with real examples
OWASP lists prompt injection as LLM01. See the direct and indirect types, documented examples and the layered defenses that reduce the damage.
Oct 18, 20264 min read
AI Security
What should you log and alert on for prompt injection (LLM01)?
Log inputs with their source, tool calls with arguments, approvals and refusals. Alert on tool use that does not match the task. A defensive LLM01 list.
Oct 18, 20264 min read
AI Security
LLM excessive agency: what to log and alert on
OWASP LLM06 warns about agents with too many tools, permissions or autonomy. Log every tool call and approval, and alert on scope, volume and denials.
Oct 18, 20264 min read
AI Security
Evaluating an LLM for security work: the evals that matter
Public benchmarks will not tell you if a model helps your security team. Build evals from your own tasks, score false positives and rerun on every change.
Oct 18, 20264 min read
Stay informed
Weekly insights on software supply chain security, delivered to your inbox.
Self-healing security runs on Safeguard.
Your first fix PR is minutes away.
No sales call required, even your agent can complete the purchase over MCP.