不点一下就能被黑：Black Hat 2026 揭开的

This was not an isolated bug. It was the most dramatic demonstration of a previously undocumented class of vulnerabilities Zenity calls "PleaseFix" — a family of zero-click flaws that affected every major commercial agentic browser the researchers examined .

How the Atlas WhatsApp Worm Worked

Zenity's proof-of-concept attack bypassed OpenAI's three-tier safety system using three techniques :

• Hebrew-language instructions: Atlas's safety filters were primarily trained on English, so non-English commands evaded detection .
• Intent collision: The researchers blended legitimate user commands (like "subscribe to newsletter") with hidden malicious ones, making it hard for the agent to distinguish safe from unsafe actions .
• Fake sandbox claims: The attack page told Atlas the actions were occurring in a safe testing environment, tricking the agent's own safety self-checks .

The same exploit chain could also be used to make unauthorized Amazon purchases, demonstrating a broader account-takeover capability . WhatsApp's end-to-end encryption was not broken — the AI agent was simply acting within the user's authenticated session .

The PleaseFix Vulnerability Class

Zenity Labs disclosed PleaseFix as a family of zero-click vulnerabilities affecting every major commercial agentic browser they examined . The flaws were demonstrated across:

• ChatGPT Atlas (OpenAI)
• Claude in Chrome (Anthropic)
• Gemini in Chrome (Google)
• Perplexity Comet (Perplexity)
• Copilot Edge (Microsoft)

In total, Zenity found 20 distinct flaws across these platforms . The core issue is architectural: agentic browsers are designed to autonomously read web content, follow instructions, and perform actions on the user's behalf — and that autonomy is itself the attack surface . Attackers hijack agents through ordinary content and expected actions with no malware, no exploits, and no user click required .

The demonstrated attack outcomes included :

• Silent data theft: exfiltration of local files and credentials, including from password managers like 1Password .
• Account and device takeover: full session hijacking and remote code execution on the machine running the browser .
• Social account hijacking: posting, messaging contacts, and spreading phishing links from the victim's accounts.
• Unauthorized financial actions: making purchases on Amazon and other services.

Industry Responses

• OpenAI had already announced the discontinuation of Atlas before Black Hat 2026, with the browser shutting down on August 9, 2026 — just days after the revelations . OpenAI moved Atlas's agent features into its broader ChatGPT Work and Codex platforms, though security questions followed the agent capabilities to those new environments .
• Multiple vendors received coordinated disclosures from Zenity and released patches, but the fundamental architectural tension — granting AI agents broad autonomy to act on the web — remained unresolved .

Why Experts Are Calling for Deterministic Security Barriers

Security researchers argue that the current approach — layering safety filters and prompt-level guardrails on top of highly autonomous AI agents — is structurally inadequate . Zenity's research showed that every single agentic browser could be compromised using the same fundamental technique: feeding the agent malicious content it was designed to process .

The core problem is that probabilistic AI safety filters (which guess whether an action is safe) cannot reliably stop an attacker who can craft novel inputs that slip past those filters. Researchers are calling for deterministic security barriers — hard, enforceable constraints on what an AI agent is allowed to do, regardless of how it interprets instructions . Examples include:

• Hardcoded allow-lists of permitted actions (e.g., "never send messages from this browser").
• Operating-system-level permission boundaries that cannot be overridden by model output.
• Human-in-the-loop gates for high-risk actions (financial transactions, messaging, credential access).

Until such deterministic controls exist, the PleaseFix class of flaws shows that agentic browsers can be weaponized against their own users simply by publishing ordinary-looking web content .