Web LLM Prompt Injection
Modern web apps embed LLM features (chatbots, support assistants, summarisers, agents with tools). Prompt injection makes the model treat attacker-controlled text as instructions, overriding the developer's system prompt. Because the LLM often has access to APIs, user data, and…
Why it matters
Modern web apps embed LLM features (chatbots, support assistants, summarisers, agents with tools). Prompt injection makes the model treat attacker-controlled text as instructions, overriding the developer's system prompt. Because the LLM often has access to APIs, user data, and tools, this becomes a web vulnerability: data exfiltration, IDOR-style access to other users, SSRF, and even XSS when model output is rendered unsanitised.
How we test it
Our operators follow a structured checklist for this topic, adapted to your API surface and authentication model.
Testing approach
- Structured validation of Web LLM Prompt Injection during scoped assessment.
- Evidence captured with reproducible steps for your engineering team.
What we look for
- Misconfigurations and control gaps related to Web LLM Prompt Injection.
- Exploitable paths that could affect confidentiality, integrity, or availability.
Related topics
| Topic | Summary |
|---|---|
| Application Logic Flaws | Application Logic Flaws occur when a web application's normal logic or workflow can be manipulated, allowing attackers to bypass security measures or misuse functionality in unexpected ways. |
| Arbitrary Cookie Flags | Cookies that hold session or sensitive data are set without the proper security flags (Secure, HttpOnly, SameSite). This lets cookies travel over cleartext, be read by JavaScript, or be sent on cross-site requests, enabling theft and CSRF. |
| Back Button Attack | After a user logs out (or on a shared device), pressing the browser Back button or reopening cached pages reveals sensitive data or restores an authenticated view. This happens when the server lets responses be cached and does not fully invalidate the session. |
| Broken Authentication | Flaws in how an application confirms a user's identity, letting attackers compromise passwords, tokens, or sessions and assume other users' identities. It is one of the highest-impact web risks (OWASP Top 10: Identification and Authentication Failures). |
| Brute Force Attack | An attacker systematically tries many credentials, tokens, or values until one works. It succeeds when an endpoint lacks rate limiting, lockout, or sufficient secret entropy. |
| Captcha Bypass | A CAPTCHA is meant to block automated requests, but flawed implementation lets attackers skip or solve it programmatically. Bypassing it re-enables brute force, spam, and other automation the control was meant to stop. |