LLM hallucinations and their security impact
How LLM hallucinations create security risks from financial liability to supply chain attacks via slopsquatting, with mitigation strategies at every layer.
Exfiltration attacks through LLM output
How Markdown image rendering in LLM applications enables data exfiltration, covering the mechanism, indirect prompt injection delivery, and real-world CVEs.
Insecure function calling in LLM applications
How function calling in LLM applications creates code execution, excessive agency, and injection risks, with techniques for each vulnerability class.
Command injection through LLM output
How command injection arises when LLMs translate natural language into shell commands, covering unrestricted execution, guardrail bypasses, and non-determinism.
SQL injection through LLM output
How SQL injection arises in text-to-SQL LLM applications, covering data exfiltration, UNION-based guardrail bypass, and data manipulation through query types.
Cross-site scripting through LLM output
How cross-site scripting arises when LLM output bypasses HTML encoding, covering reflected XSS, stored XSS, and the external script resilience bypass.
Insecure output handling in LLM applications
LLM output is untrusted data. This article covers insecure output handling, the injection risks it enables, and where it sits in OWASP LLM05 and Google SAIF.
Prompt injection mitigations and defences
No single mitigation eliminates prompt injection. Covers prompt engineering, filtering, fine-tuning, adversarial training, and guardrail models.
LLM vulnerability scanning with garak
Garak is an open-source LLM vulnerability scanner that automates adversarial testing for prompt injection, jailbreaks, and encoding bypasses. Full walkthrough.
Introduction to LLM jailbreaking
LLM jailbreaking bypasses safety alignment to force models into generating restricted content. Covers DAN, roleplay, token smuggling, and adversarial suffixes.