AI Security
In-depth guides and analysis on ai security from the Safeguard engineering team.
676 articles
The Onboarding Tax: Why Signup Forms Break Agent Workflows
Every signup form, verification email, and OAuth redirect is a wall an AI agent can't climb. Zero-touch onboarding lets an agent create the account and sign in itself — no browser, no human relay.
Agentic Commerce: Why Your SaaS Has to Let AI Agents Buy
AI agents already research, compare, and recommend software — but the moment they hit a paywall, they stall and hand the job back to a human. Here's why that gap is expensive, and how agent-native purchasing closes it.
AI agents in AppSec pipelines: triage, remediation, and guardrails
GitHub's Copilot Autofix cuts median fix time from 1.5 hours to 28 minutes — but a 2025 Replit agent incident shows why autonomy needs hard limits.
AI risk management best practices: a lifecycle framework
NIST's AI RMF has four functions and MITRE ATLAS now tracks 84 adversarial techniques — most AI risk programs still only cover one lifecycle stage.
When Shared AI Chats Become Public Search Results
In July 2025, ~4,500 shared ChatGPT conversations turned up indexed on Google. Here's why sharable AI chats leak, and how enterprises stop it at the gateway.
A risk framework for enterprise AI coding and agent tool rollouts
Samsung banned ChatGPT company-wide after three leaks in 20 days. A working framework for data exposure, model supply chain, and access control risk.
Why traditional AppSec still catches most enterprise AI agent bugs
OWASP's LLM Top 10 names new categories, but most enterprise agent breaches trace back to broken access control and unvalidated input — the classics.
Concrete guardrails for AI coding assistants
40% of Copilot-generated code contained CWE Top 25 flaws in a 2022 study. Here are the prompt, scanning, and review gates that actually stop AI-written risk.
The guardrail gap in low-code agentic AI platforms
Low-code AI builders let business users wire agents to live connectors in minutes — but most ship without per-tool scoping, approval gates, or audit trails.
Least-Privilege Tool Scoping for AI Coding Agents
One overprivileged GitHub token let researchers hijack an AI agent into leaking private repo data via a public issue. Scoping tool access closes that gap.
MCP server security for AI coding agents
A critical RCE in Anthropic's own MCP Inspector (CVSS 9.4) and two Cursor CVEs show that giving AI agents tool access creates a new, largely unvetted attack surface.
A Reproducible Rubric for Measuring Prompt-Injection Risk in Agent Skills
OWASP has ranked prompt injection the #1 LLM risk for two straight editions, yet almost no one scores agent skill packages for it consistently. Here's a rubric.