Accessibility-tree injection safety
Instruction-shaped text in aria-labels and alt text is read by an agent as part of the page, and can be used to steer it.
- Category
- Agent experience
- Standard
- Recommended
Who supports it
ClaudeSupported (announced)
Published 25 Aug 2025 · last checked 9 Oct 2026: Claude blog: Piloting Claude in ChromeClaude in Chrome reads DOM content invisible to humans: 'hidden malicious form fields in a webpage's Document Object Model (DOM) invisible to humans ... URL text and tab title that only an agent might see.' Post updated through 2025-12-18. See also Use Claude in Chrome safely (support.claude.com/en/articles/12902428) and the browser use tool's injection classifiers.
CometPartial support
Published 25 Nov 2025 · last checked 9 Oct 2026: BrowseSafe: Understanding and Preventing Prompt Injection Within AI Browser Agents (Perplexity / Purdue, arXiv 2511.20597)Perplexity co-authored paper; benchmark built from Comet agent tool outputs. Hidden-injection strategies include 'HTML comment, data attribute, CSS-hidden text, form hidden fields, and semantic attribute abuse'. Not an explicit statement that aria-label/alt are passed to Comet.
Gemini in ChromeSupported (announced)
Published 8 Dec 2025 · last checked 9 Oct 2026: Google blog: Architecting security for agentic capabilities in ChromeIndirect prompt injection 'can appear in malicious sites, third-party content in iframes, or from user-generated content'; Gemini plans from 'page content shared in Chrome'. Defences: User Alignment Critic, origin gating, confirmations. Does not say which attributes are included.
24 others: no evidence yet. See every agent and browser
What it checks
The extension reads the aria-label, alt, title and placeholder
attributes on the rendered page and looks for text shaped like an instruction to
an AI model. It matches phrases such as:
- “ignore previous instructions” (or “prior”, “above”, “all previous”)
- “disregard the above” or “disregard previous”
- “system prompt”
- “you are now …”
- “new instructions:”
This check deliberately includes elements a person cannot see, such as
zero-opacity or zero-size text, because that is how an instruction is hidden
from people while staying readable to an agent. It only skips elements removed
from the accessibility tree altogether (hidden, inert, aria-hidden="true",
display: none or visibility: hidden), since an agent never sees those either.
Results
| Status | When |
|---|---|
| Pass | No instruction-shaped text was found |
| Fail | One or more attributes contain instruction-shaped text |
| Unmeasured | The extension could not read the page (open it in the active tab) |
The result quotes each matching attribute.
How to fix
Remove command-style text aimed at a reader from aria-label, alt, title
and placeholder. These attributes should describe the element, not address
whoever is reading the page.
If you did not put the text there, treat it as a security problem: look for injected content from user submissions, third-party widgets or a compromised plugin. Agents and the tools that protect them may flag the page as a prompt injection attempt.
Support by agent and browser
Announced means the vendor’s own docs or announcement say so. Published is the date on the vendor’s own page, where it has one. Checked is the last time our weekly research found that page still saying so.
| Agent | Status | Source |
|---|---|---|
| AI assistants | ||
| Claude Anthropic | Supported (announced) | Claude blog: Piloting Claude in Chrome Published · Checked Claude in Chrome reads DOM content invisible to humans: 'hidden malicious form fields in a webpage's Document Object Model (DOM) invisible to humans ... URL text and tab title that only an agent might see.' Post updated through 2025-12-18. See also Use Claude in Chrome safely (support.claude.com/en/articles/12902428) and the browser use tool's injection classifiers. |
| AI browsers | ||
| Comet Perplexity | Partial support | partial BrowseSafe: Understanding and Preventing Prompt Injection Within AI Browser Agents (Perplexity / Purdue, arXiv 2511.20597) Published · Checked Perplexity co-authored paper; benchmark built from Comet agent tool outputs. Hidden-injection strategies include 'HTML comment, data attribute, CSS-hidden text, form hidden fields, and semantic attribute abuse'. Not an explicit statement that aria-label/alt are passed to Comet. |
| Gemini in Chrome Google | Supported (announced) | Google blog: Architecting security for agentic capabilities in Chrome Published · Checked Indirect prompt injection 'can appear in malicious sites, third-party content in iframes, or from user-generated content'; Gemini plans from 'page content shared in Chrome'. Defences: User Alignment Critic, origin gating, confirmations. Does not say which attributes are included. |
No evidence yet24
- ChatGPT
- Gemini
- Perplexity
- Copilot
- Meta AI
- ChatGPT Atlas
- Dia
- Edge Copilot Mode
- Claude Code
- Cursor
- Codex
- GitHub Copilot
- Windsurf
- CCBot
- Applebot
- Amazonbot
- Bytespider
- Cohere
- Mistral
- DuckAssist
- Chrome
- Edge
- Firefox
- Safari
Names and marks belong to their owners and are used only to identify products. None of these vendors is affiliated with Sitebulb. Spotted something out of date? Send a correction.
Adoption on the top 100 sites
100 of 100 measured sites implement it
Sites implementing it100
- #1 google.com
- #2 cloudflare.com
- #3 aws.amazon.com (via amazonaws.com)
- #4 facebook.com
- #5 gstatic.com
- #8 googleapis.com
- #9 microsoft.com
- #11 youtube.com
- #12 apple.com
- #13 instagram.com
- #15 mail.ru
- #16 x.com (via twitter.com)
- #18 linkedin.com
- #20 dzen.ru
- #23 googletagmanager.com
- #24 azure.microsoft.com (via azure.com)
- #26 m365.cloud.microsoft (via office.com)
- #27 outlook.live.com (via live.com)
- #29 amazon.com
- #32 github.com
- #33 wikipedia.org
- #37 fastly.com (via fastly.net)
- #39 whatsapp.com (via whatsapp.net)
- #42 bing.com
- #43 netflix.com
- #44 triplinkintl.com
- #45 googleusercontent.com
- #48 gandi.net
- #50 marketingplatform.google.com (via doubleclick.net)
- #52 uk.pinterest.com (via pinterest.com)
- #53 icloud.com
- #55 wordpress.org
- #56 teams.live.com (via skype.com)
- #57 cloudflare.net
- #59 goo.gl
- #60 digicert.com
- #61 adobe.com
- #63 tiktok.com
- #67 domains.google (via googledomains.com)
- #68 zoom.com (via zoom.us)
- #69 uk.yahoo.com (via yahoo.com)
- #70 sso.myfritz.net (via myfritz.net)
- #71 endpoints.news
- #72 api.whatsapp.com (via wa.me)
- #75 roblox.com
- #76 opera.com
- #77 open.spotify.com (via spotify.com)
- #80 baidu.com
- #81 tiktokv.com
- #83 ya.ru (via yandex.net)
- #84 nginx.org
- #85 ntp.org
- #86 msn.com
- #87 qq.com
- #90 vimeo.com
- #91 primevideo.com (via amazonvideo.com)
- #93 bunny.net (via b-cdn.net)
- #94 blogger.com (via blogspot.com)
- #95 mozilla.org
- #96 f5.com
- #98 telegram.org (via t.me)
- #99 store.epicgames.com (via epicgames.com)
- #101 wordpress.com
- #103 reddit.com
- #106 sentry.io
- #107 letsencrypt.org (via lencr.org)
- #109 ui.com
- #110 european-union.europa.eu (via europa.eu)
- #112 chatgpt.com
- #114 st.okcdn.ru (via okcdn.ru)
- #116 bitly.com (via bit.ly)
- #117 ipify.org
- #119 nic.ru
- #120 samsung.com
- #121 office365.com
- #122 intuit.com
- #123 aliyuncs.com
- #127 openai.com
- #128 one.one.one.one (via cloudflare-dns.com)
- #130 registrar-servers.com
- #131 pages.github.com (via github.io)
- #132 dropbox.com
- #133 skybridge.click
- #135 discord.com (via discord.gg)
- #136 developer.adobe.com (via adobe.io)
- #139 apache.org
- #140 android.com
- #141 nih.gov
- #144 vk.ru
- #147 reg.ru
- #150 t-online.de
- #151 archive.org
- #152 forms.gle
- #153 webex.com
- #154 snapchat.com
- #156 godaddy.com
- #158 kaspersky.co.uk (via kaspersky.com)
- #159 firebase.google.com (via crashlytics.com)
- #160 tumblr.com
- #161 nytimes.com
Overview
How this rating is worked out5 of 9 points
| Factor | Points | Why |
|---|---|---|
| Vendor support | 2 / 3 | 2 vendors supporting, 1 partly or planned |
| Top-100 adoption | 3 / 3 | 100% of measured sites |
| Spec maturity | 0 / 3 | Community proposal |
8 or more points is Established, 5 to 7 is Gaining traction, 2 to 4 is Speculative and under 2 is Stalled. A feature is also Stalled when more vendors have said no than yes. Read the spec.
Community insights
No community insights for this feature yet.
Quotes are from public posts, checked against the original before they appear here. Suggest an insight, or ask for yours to be removed.