AI Security
In-depth guides and analysis on ai security from the Safeguard engineering team.
786 articles
How to build an AI-specific incident response playbook
A step-by-step guide to building an AI incident response plan — covering scoping, escalation, detection, containment, and post-incident review for LLM and agent failures.
CoSAI Releases Model Signing and Incident Response Frameworks
The Coalition for Secure AI published two operational frameworks in November 2025: Signing ML Artifacts and AI Incident Response. We unpack what each contains and how to adopt them.
Prompt Injection Detection in Retrieval Systems
Indirect prompt injection arrives through your retrieval corpus, not your chat box. We cover the detection strategies that survive when attackers write your RAG content.
The Hugging Face Breach: What Changes When an AI Agent Runs the Intrusion
On 16 July 2026 Hugging Face disclosed that a malicious dataset gave an attacker code execution inside its data-processing pipeline, escalating to node-level access and internal cluster credentials over a single weekend — driven by an autonomous agent framework executing thousands of actions. Here's the anatomy, and what it means for anyone who treats a model registry as a trusted input.
vLLM CVE-2025-62164: Tensor Deserialization RCE
vLLM 0.10.2-0.11.0 deserialized user-supplied PyTorch tensors via torch.load() in the Completions API. Memory corruption, potential RCE.
Claude Opus 4.5 System Card: Defender Takeaways
Anthropic released Claude Opus 4.5 on November 24, 2025 with the most detailed safety section of any system card to date. We pull out what enterprise defenders should change.
AI Model Watermarking and Provenance
Watermarking and provenance are the two most confused terms in AI security. A practical breakdown of what each actually does, where the 2025 techniques break, and what to ship in the meantime.
vLLM CVE-2025-66448: Auto-Map RCE via Model Configs
A critical RCE in vLLM allows malicious model configs to bypass trust_remote_code=False. We analyze the bug, the patch, and what every vLLM operator should do.
Gemini 3 Pro and the Frontier Safety Framework Report
Google released Gemini 3 Pro on November 18, 2025 with the most thorough Frontier Safety Framework evaluation yet. We unpack what was disclosed and how it changes downstream defender posture.
Anthropic's Responsible Scaling Policy v3: What Changed
RSP v3.0 takes effect February 24, 2026. It splits the AI R&D threshold, adds a CBRN-development tier, and formalizes Risk Reports.
ShadowMQ: 30+ RCE Flaws Across AI Inference Engines
Oligo Security disclosed ShadowMQ in November 2025: ZeroMQ-and-pickle deserialization patterns copy-pasted across vLLM, Meta Llama, TensorRT-LLM, and others. We dissect the pattern.
Training Data Provenance for Enterprise Fine-Tuning
Fine-tuning corpora are supply chain artifacts. We cover the provenance signals, attestations, and drift controls enterprises need before pushing weights to prod.
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.