Building secure AI agents
Vercel discusses security risks in AI agents, highlighting prompt injection as a critical vulnerability where attackers can manipulate model behavior through controlled inputs. Developers building agents must design with adversarial scenarios in mind, treating user input and external data as potentially untrusted.
An AI agent is a language model with a system prompt and a set of tools. Tools extend the model's capabilities by adding access to APIs, file systems, and external services. But they also create new paths for things to go wrong.
The most critical security risk is . Similar to SQL injection, it allows attackers to slip commands into what looks like normal input. The difference is that with LLMs, there is no standard way to isolate or escape input. Anything the model sees, including user input, search results, or retrieved documents, can override the system prompt or event trigger tool calls.prompt injection
If you are building an agent, you must design for worst case scenarios. The model will see everything an attacker can control. And it might do exactly what they want.
Source: original entry ↗
More from Vercel
Follow Vercel to get its new changes in your feed and email digest.
OpenAI Decisions API now available on AI Gateway
OpenAI's Decisions API is now accessible through Vercel's AI Gateway with an OpenAI-compatible endpoint, enabling decision models to answer typed questions and return probabilities, choices, and scores for routing, triage, and guardrails use cases. Support is available across the OpenAI SDK, AI SDK, HTTP API, and CLI with the latest versions.
Timestamp attributes now supported in Vercel Flags
Vercel Flags now supports timestamp attributes for entities, allowing you to create time-based targeting rules. Use this feature to run limited-time campaigns, show content between specific dates, or target users based on registration date.
Glyph Cluster now available in stealth on AI Gateway
Glyph Cluster, a reasoning model for coding and long-context analysis, is now available as a stealth model on Vercel's AI Gateway for Pro and Enterprise plan teams with purchased AI Gateway credits at no cost during the stealth period. The model supports function calling, streams responses, and can be accessed via AI SDK, OpenAI-compatible APIs, and coding agents.