Guide from the official docs
Prompt injection: can a web page or email hijack your AI agent?
What OWASP, NIST and the UK NCSC say prompt injection is, what each agent vendor documents about it, published attack rates and the steps vendors advise.
This guide follows the official docs and announcements, linked inline. AideDex did not test these products and does not score them. How we check facts.
Last checked Oct 24, 2026.
Yes. A web page, email or document can carry hidden instructions that steer an AI agent away from your request, and no vendor below claims to have stopped this fully. Every one of them says it lowers the risk with classifiers, confirmations or blocked sites. They also ask you to limit what the agent can reach and to watch what it does.
AideDex has not tested these products. This guide reports only what standards bodies and vendors publish. For email access and logins in particular, see Is it safe to connect Gmail to an AI agent? and AI agents, passwords and payments.
What prompt injection is
OWASP ranks prompt injection first (LLM01) in its 2025 Top 10 for LLM applications. It separates direct injection, typed by the user, from indirect injection, where content from “external sources, such as websites or files” changes how the model behaves. The instructions “do not need to be human-visible/readable”, as long as the model parses them (OWASP). OWASP also says it is “unclear if there are fool-proof methods of prevention”. Its mitigations include least-privilege access, human approval for high-risk actions, and marking untrusted content (OWASP).
NIST calls the agent version agent hijacking: “a type of indirect prompt injection” where an attacker plants instructions in an email, file or website that the agent later reads (NIST).
The UK’s National Cyber Security Centre (NCSC) says current models “simply do not enforce a security boundary between instructions and data inside a prompt”. So prompt injection “may never be totally mitigated” the way SQL injection can be. Its advice is to design for reducing likelihood and impact, and to be wary of any product claiming to “stop” it (NCSC).
Published attack-success numbers
Few vendors publish a rate. These are the figures as published, from different tests that cannot be compared directly:
| Who published it | What was measured | Figure |
|---|---|---|
| Anthropic | Claude in Chrome’s current configuration against Anthropic’s internal tests combining known effective attacks | “less than 0.08%” attack success (Anthropic) |
| NIST (CAISI) | Agents powered by Claude 3.5 Sonnet (Oct 2024 upgrade) in the AgentDojo test framework | In AgentDojo’s Workspace environment, 11% for the strongest baseline attack, 81% for the strongest new red-team attack. Attempting each of five injection tasks 25 times raised the average success rate from 57% to 80% (NIST) |
| OpenClaw docs, citing a 2026 crowdsourced arena | 272K attacks across 41 agent scenarios, counted only when the agent did the harmful action and hid it | 0.5% Claude Opus 4.5, 1.0% Sonnet 4.5, 1.3% Haiku 4.5, 8.5% Gemini 2.5 Pro. The same docs cite published success rates above 80% against state-of-the-art defenses once the attacker adapts (OpenClaw) |
OpenAI, Google, Microsoft, Perplexity, Opera and xAI do not publish a success rate on the pages cited below.
What each vendor documents
ChatGPT Work (cloud browser) and Atlas
- OpenAI says it “tests for risks such as prompt injection, phishing, and unintended actions”, and that those safeguards “do not eliminate every risk” (OpenAI Help).
- By default ChatGPT asks before visiting a new site. Under Settings > Cloud browser you can pick Always ask, Auto approve or Always allow, and OpenAI says Always allow “is not recommended”. It asks for confirmation before actions such as a booking or payment (OpenAI Help).
- For Atlas agent mode, OpenAI says prompt injection is “unlikely to ever be fully ‘solved’”. It recommends logged-out mode when sign-in isn’t needed, reviewing confirmation requests and avoiding broad prompts (OpenAI).
ChatGPT dots
- OpenAI says a web page, email or document can carry malicious instructions. It combines “model safeguards with tool restrictions, checks before actions, and monitoring” (OpenAI).
- A separate Auto-review system checks planned actions such as sending email against your instructions. Background research uses read-only tools that “cannot directly send messages” (OpenAI). Plans and status: ChatGPT dots.
Claude in Chrome and Claude Cowork
- Anthropic calls prompt injection “the biggest risk facing browser-using AI tools”. One classifier checks incoming content and another checks every action, and flagged actions are blocked or paused. “The risk is not zero” (Anthropic).
- For Cowork, Anthropic says an attack needs two things at once: Claude can read content from outside your trusted sources, and it can take harmful actions. It advises switching to “Manually approve” for sensitive or hard-to-undo work (Anthropic). Agent pages: Claude in Chrome and Claude Cowork.
Perplexity Comet
- Perplexity describes four layers: classifiers that scan retrieved content (hidden HTML/CSS text, image-based injection, goal hijacking), prompts that mark web content as untrusted, confirmation before sending emails, changing your calendar or placing orders, and notifications when something is blocked (Perplexity).
- It calls malicious prompt injection “an unsolved problem across the industry” (Perplexity), and says it hired Trail of Bits to audit Comet before launch (Perplexity).
- In Settings → Privacy and security you can block tasks on specific websites, such as your bank (Perplexity Help). More on the browser: Perplexity Comet, and for first-run settings and the agent permission prompt, how to get started with Perplexity Comet.
Gemini tasks
- Google says malicious instructions may be “hidden from you but visible to the AI agent”. Its examples include posting your private info publicly and sending your Gmail to an outside service (Google).
- Google says Gemini asks for confirmation before sending messages, deleting data, buying, and submitting forms. It also limits browsing to sites relevant to the task. Google adds that these safeguards “don’t guarantee protection against all risks” (Google). Plans and regions: Gemini tasks. The browsing settings are walked through in how to get started with Gemini tasks.
Microsoft: Copilot in Edge and Copilot Actions on Windows
- For Browse with Copilot, Microsoft gives the example of hidden text on a recipe page that sends Copilot to collect sensitive information. While browsing, Copilot can’t use autofill data, saved passwords or wallet information. Microsoft advises against banking, trading and ID data (Microsoft).
- On Windows, Microsoft names cross-prompt injection (XPIA), where content in “UI elements or documents can override agent instructions”. Agents run under their own account in a separate session (Microsoft).
Opera Neon
- Opera says Neon analyzes prompts for malicious traits, pauses for transactions and downloads, and blocks high-risk pages such as banking sites. It says the risk “cannot be entirely reduced to zero” (Opera). Brave’s October 2025 report of a hidden-text attack on Neon, and Opera’s fix, are in how to get started with Opera Neon.
Grok Bot
- xAI’s approvals page does not use the term “prompt injection”. It describes Auto Review (Settings → General → Auto-review), which checks actions before they run, with Ask first rules that always stop for you. xAI says Auto Review “is model-based and should complement, not replace, least privilege” (xAI Docs). Plans and status: Grok Bot. Setting limits before it acts is a step in how to get started with Grok Bot.
OpenClaw (self-hosted)
- The docs say any untrusted content can carry instructions, “even if only you can message the bot”. They advise sandboxing, keeping
web_search,web_fetchandbrowseroff unless needed, and not using older or smaller models for tool-enabled agents (OpenClaw). Install steps with those settings: OpenClaw setup, local and VPS. The security warnings Chinese regulators issued about OpenClaw in 2026 are in OpenClaw in China.
Steps vendors recommend
These steps come up across several vendors’ docs:
- Keep requests narrow. OpenAI warns against prompts like “review my emails and take whatever action is needed” (OpenAI).
- Keep sensitive accounts out of reach. Anthropic suggests a separate browser profile without banking, healthcare or government accounts (Anthropic). Comet lets you block sites (Perplexity Help). For bank logins and one-time codes, see AI agents, banking and 2FA.
- Read confirmation prompts before approving. Grok Bot’s docs say not to approve an action whose target or effect you cannot identify (xAI Docs).
- Stop the task if it drifts. Anthropic lists unrelated topics, unexpected sites or requests for sensitive information as possible signs of injection (Anthropic).
- Be careful with scheduled runs. Google notes that you may not be able to stop an unintended action while you are offline (Google). Anthropic advises against scheduling tasks that send messages or make purchases (Anthropic).
Some sites also block agents outright. See which websites block AI agents.
Sources
17 pages- OWASP: LLM01:2025 Prompt Injection genai.owasp.org
- NIST: Technical Blog: Strengthening AI Agent Hijacking Evaluations nist.gov
- UK NCSC: Prompt injection is not SQL injection (it may be worse) ncsc.gov.uk
- OpenAI Help Center: Using cloud browser in ChatGPT help.openai.com
- OpenAI: How we build safety, security, and privacy into dots openai.com
- OpenAI: Continuously hardening ChatGPT Atlas against prompt injection openai.com
- Anthropic Help Center: Use Claude in Chrome safely support.claude.com
- Anthropic Help Center: Use Claude Cowork safely support.claude.com
- Perplexity: Mitigating prompt injection in Comet perplexity.ai
- Perplexity: How we built security into Comet from day one perplexity.ai
- Perplexity Help Center: Control what Comet Assistant can use perplexity.ai
- Google Gemini Help: Manage your tasks & workflows in Gemini Apps support.google.com
- Microsoft Support: Browse with Copilot in Microsoft Edge support.microsoft.com
- Microsoft Support: Experimental Agentic Features support.microsoft.com
- Opera Neon: Understanding agentic browser security operaneon.com
- xAI Docs: Grok Bot approvals, security, and privacy docs.x.ai
- OpenClaw Docs: Prompt injection docs.openclaw.ai