Topics

This week's digest: SSRF through agent tool-use, the model supply-chain class and why safetensors matters, and model extraction as a business risk.

Google says it caught an attacker using an LLM to find a zero-day, peer-reviewed research shows reasoning models can autonomously jailbreak other models

A recurring AI-security finding: an injected instruction makes the model emit a markdown image whose URL carries the user's data to an attacker server.

This week's digest: indirect injection becomes the agent-era default, the markdown-rendering data-exfiltration class, and why system-prompt secrecy keeps

A pre-auth SQL injection in the LiteLLM proxy landed in CISA's KEV catalog and was exploited within roughly 36 hours of disclosure.

As LLM agents gained tools and memory, the dangerous injection stopped coming from the user and started coming from the data the agent reads.

Prompt injection tops the OWASP LLM Top 10, and the tool market has consolidated fast. A field guide to Llama Prompt Guard 2, Azure Prompt Shields, Lakera Guard, NeMo Guardrails, and the open-source options that survived.

OWASP's Q1 2026 GenAI Exploit Round-up confirms AI agent security risks have moved from research to active production breaches.

A practitioner breakdown of the OWASP LLM Top 10 2026 changes — two new threat categories dropped, three proposed additions for 2026, and a companion

A practitioner-level breakdown of LLM prompt injection attack examples — from basic instruction overrides to CVE-rated zero-click exploits in production

A technical breakdown of LLM jailbreak defense techniques, from input filtering and model-level hardening to multi-agent architectures and OWASP mitigations.

A working guide to evaluating a weekly AI security newsletter: what belongs in one, which primary sources it should cite, and how to spot vendor noise.

AI Sec Weekly is a Friday roundup of AI security news. Here's the case for a seven-day cadence, the bar an item has to clear, and how each entry is written.