Adobe patched CVE-2026-48294, a flaw in Adobe Acrobat Chrome extension that could let attackers steal WhatsApp Web chats by luring users to a webpage.Guardio Labs researcher Shaked Biner disclosed HermeticReader, a vulnerability chain in the Adobe Acrobat Chrome extension that allowed any attacker-controlled webpage to silently steal a visitor’s WhatsApp chats, contacts, profile name, and message previews in plain text. The extension sits on roughly 329 million browsers. No malware, no phished password, no compromised session cookie — just visiting the wrong page was enough.“Read the prerequisites list of this exploit chain and the discomfort sets in: no malware is installed, no password is phished, no session cookie is touched. There is no zero-day in WhatsApp.” reads the report published by Guardio Labs. “The attacker needs no Adobe account and no foothold on the machine. Only for the victim to visit a simple attacker-controlled static page.”The attack chains three separate flaws in the extension’s internal messaging system. None of them are dramatic on their own. Together they hand an attacker full DOM control over an open WhatsApp Web tab.The first flaw is that the extension’s web-accessible resource pages, frame.html, searchWidget.html, and others, can be embedded as hidden iframes by any website. Those pages parse a JSON blob directly from their URL query string and relay it to the extension’s service worker as a message. “Because that script runs inside chrome-extension://efaidn…/, the service worker sees its message as coming from a trusted internal source. The extension’s own identity becomes the attacker’s mask. The page can set any fields on the message , type, key, value, anything else.” continues the report. “Two fields are not free: panel_op has to be set to "load-frictionless" so that frame.js takes the relay code path at all, and main_op is overwritten to "relay_to_content" on the way out.”The service worker’s message listener never checks the sender. Its storage writer has no allowlist on what can be written. Any page can write any key into the extension’s local storage in zero clicks.That storage write matters because of the third flaw: Adobe’s Hermes engine, the component that handles WhatsApp Web integration, activates based on a feature flag read from local storage. Write the right key – floodgate-add = "dc-cv-hermes" – and Hermes arms immediately. With Hermes running, the attacker controls a command dispatcher inside WhatsApp’s tab that can inject HTML, invoke methods on any DOM node, submit forms, and relay messages back to the service worker. One of those commands, APPEND_HTML_TO_TARGET, can replace WhatsApp’s login QR code with an attacker-controlled one, meaning anyone who scans it to link a new device pairs the attacker’s session instead.The data exfiltration technique is what makes the research genuinely clever. “We never asked the extension to read anything. We asked WhatsApp to submit itself to us – and it did.” states the report.The attack injects a POST form into WhatsApp’s DOM, then uses the ELEMENT_OPERATION command to physically move WhatsApp’s entire body node into the form’s option element. Because an option element with no value attribute submits its text content, and because WhatsApp’s content security policy contains no form-action directive, submitting the form sends the entire rendered page, chat list, contact names, message text, to the attacker’s server. WhatsApp does the exfiltration itself.Guardio found this within hours of Adobe shipping version 26.5.2.1 on June 3, using a custom agentic AI system that unpacked and analyzed the extension’s 344 obfuscated JavaScript files, mapped code diffs against previously analyzed flows, and worked through a 138-case service-worker message dispatcher. “The Agentic AI Research harness is not only a clever and innovative way to research – it already became mandatory! From a new version release trigger, the work is shared: The agent unpacks the bundle, beautifies 344 obfuscated JavaScript files, finds the code diffs and maps them against already mapped and analyzed flows, continues to map a 138-case service-worker message dispatcher, and so much more.” states Guardio Labs. “At this time, we’ve steered the analysis toward what is actually reachable and worth proving, cleared dead ends, and presented plausible ideas and attack vectors to pursue. That combination is what collapsed the time line into hours instead of days, weeks or even months.”Adobe’s response matched the speed: acknowledged, patched, and shipped within the same weekend the report arrived. CVE-2026-48294 was issued days later.The structural lesson here isn’t about any single clever trick. Twelve individually unremarkable shortcuts in message passing, storage handling, feature flags, and host matching composed into a chain that reached 329 million browsers. The era in which a high-install extension could rely on nobody looking at the plumbing closely is ending, for defenders and attackers alike.Follow me on Twitter: @securityaffairs and Facebook and MastodonPierluigi Paganini(SecurityAffairs – hacking, Adobe Acrobat)