New Webmail Exploits Bypass Security Boundaries, Risking Data Theft and Account Takeovers Research presented at Black Hat USA 2026 by PortSwigger’s Gareth Heyes reveals critical vulnerabilities in major webmail platforms including Outlook, Gmail, Fastmail, Proton Mail, Yahoo Mail, and AOL Mail that allow malicious content within emails to escape message boundaries and manipulate trusted interfaces. The findings, published on August 6, demonstrate how attackers could steal passwords, hijack third-party accounts, leak tokens, and exploit AI tools processing emails. ### Key Attack Vectors 1. Outlook (Firefox Chain) - A proof-of-concept (PoC) spoofs a Microsoft sign-in screen, capturing passwords as users type. The attack leverages label elements, custom attributes, and media-query parsing tricks to bypass sanitization, disguising a select element as a password field. Firefox’s option-selection timer reset enables real-time credential theft. 2. Yahoo/AOL (Paste Race Condition) - In Firefox, pasted HTML retains active CSS briefly before sanitization, allowing attackers to extract Medium email-login tokens. The victim copies attacker-supplied CSS into a draft, triggering requests that reveal enough of the 12-character token for reconstruction and account takeover. 3. Gmail (AI-Powered Exfiltration) - A prompt-injection attack via Anthropic’s Claude Cowork (connected to Gmail) tricks the AI into retrieving a Slack token from a confirmation email. The token is placed in an HTML draft, and viewing it leaks the data. The exploit abuses Gmail’s `image-set()` fallback to bypass sanitization. 4. Fastmail (CSS Hotwiring & AI Manipulation) - CSS pseudo-elements and opacity tricks deceive OpenAI’s Atlas AI (deprecated as of August 9, 2026) into executing hidden instructions. When a user asks Atlas to translate visible text, the AI instead opens tabs and encodes the victim’s name in URLs. - "CSS hotwiring" redirects clicks to unintended UI actions, while an image-proxy bypass (via the `user.fm` domain) reveals email-viewing activity. 5. Proton Mail (IP Leak) - A separate exploit exposes the recipient’s IP address, contradicting Proton’s claim of hiding personal IPs and exact open times. ### Defensive Gaps & Mitigations While Fastmail patched two CSS mutation bugs and a Proton Mail proxy bypass was neutralized, several vulnerabilities remain unaddressed: - Outlook’s label-jacking and Gmail’s `image-set()` bypass still function. - The full Outlook password-capture chain’s fix status is unclear. The research highlights two primary attack paths: - Abusing allowed HTML/CSS in webmail interfaces. - Exploiting discrepancies between sanitizer approvals and browser rendering. ### Recommended Defenses The paper advises webmail providers to: - Isolate HTML emails in sandboxed iframes. - Restrict CSS, custom attributes, select menus, and image requests. - Use character allow lists for CSS validation and block dangerous selectors. - Prevent attacker-controlled image requests via allow-listed domains. Public PoCs for the disclosed techniques remain available as of August 8, though no active malicious exploitation has been reported. The findings underscore the need for strict boundary enforcement between untrusted email content and trusted interfaces.

Microsoft Outlook Add-In Hijacked to Steal Thousands of Credentials and Payment Data Security researchers at Koi Security uncovered a novel attack leveraging a dormant Microsoft Outlook add-in to harvest over 4,000 login credentials, credit card numbers, and banking security answers. The incident marks the first known malicious Office add-in discovered in the wild, exposing a critical flaw in Microsoft’s third-party tool distribution. The attack centered on AgreeTo, a legitimate meeting-scheduling add-in published to the Microsoft Office Add-in Store in 2022. After the developer abandoned the project and its hosting domain expired, the subdomain (`outlook-one.vercel.app`) became available for registration. An attacker claimed the domain and replaced the original tool with a fake Microsoft sign-in page, which loaded inside Outlook via an iframe for all existing users. Microsoft’s security review process only validates an add-in’s manifest file upon initial submission, meaning the malicious content change went undetected. The phishing page captured credentials and transmitted them to the attacker via a Telegram bot. Researchers accessed the exfiltration channel, recovering stolen data including Microsoft account logins, payment details, and IP addresses while the attackers were actively testing the credentials. Though Microsoft removed the add-in from its store, the phishing infrastructure remained operational outside it. The AgreeTo manifest had ReadWriteItem permissions, granting potential access to read or modify users’ emails, though the attackers only deployed a basic phishing scheme. The incident underscores a broader vulnerability in software supply chains: Office add-ins function as remote dynamic dependencies, allowing content to change without Microsoft’s oversight.