Lakera bulletin - This week in AI

[_Val_]

It’s been a busy week for AI security and infrastructure. From stealthy browser-based prompt injections to Microsoft and OpenAI staking claims on the “AI browser” frontier, the theme is clear: as AI moves deeper into our daily workflows, its attack surface expands.

We’re also spotlighting Lakera’s 2025 GenAI Security Readiness Report, a reality check on how prepared organizations truly are for this new wave.

Let’s get into it.

AI Sidebar & Browser Extensions Can Be Abused

Researchers found that malicious browser extensions and sidebar features in AI-enhanced browsers can be used to spoof trusted AI assistants, steal credentials, and hijack browsing sessions, exposing an overlooked threat vector as browsers evolve into AI platforms.
🔗 Read the full story

Prompt Injection via Hidden Instructions Hits AI Browsers

A new attack technique hides malicious instructions inside images and screenshots to hijack AI browsers such as Perplexity Comet. The exploit can bypass sanitization filters, underscoring the need for defensive layers against invisible prompt-based attacks.
🔗 See Brave’s analysis

Zero-Day SharePoint Exploit Used by Chinese Threat Actors

A patched Microsoft SharePoint vulnerability (CVE-2025-53770) was recently exploited by multiple Chinese threat actors to breach global telecom and government networks. The campaign highlights the continuing risk of delayed patching in enterprise environments.
🔗 Read more at The Hacker News

Supply-Chain Risks Raise Red Flags After F5 Networks Breach

F5 Networks, a major cybersecurity vendor, confirmed a year-long intrusion that may have exposed enterprise and AI-infrastructure customers. Experts warn this could be one of the most consequential supply-chain breaches in years.
🔗 Read the report

DeepSeek Opens Up Text-to-Image Compression Model for Longer Context Windows

DeepSeek released an open-weight model, DeepSeek-OCR, that compresses text through image representation 10× more efficiently, paving the way for LLMs capable of tens of millions of tokens in context.
🔗 Explore the announcement

Microsoft Copilot and Edge Get Major Fall Update: The “AI Browser” Era Begins

Microsoft rolled out 12 new Copilot features, including “Copilot Mode” in Edge that transforms browsing into a fully AI-assisted workspace, with voice control, multi-tab context, and history-aware recall.
🔗 Read the Microsoft announcement

OpenAI Launches ChatGPT Atlas

OpenAI introduced ChatGPT Atlas, a new AI-powered web browser built around ChatGPT. Atlas combines conversational browsing, “Agent Mode” for task automation, and optional on-device memory, now available for macOS, with Windows, iOS, and Android on the way.
🔗 Read the official announcement

ICYMI: Lakera’s 2025 GenAI Security Readiness Report

Lakera’s new report offers one of the most comprehensive looks yet at how organizations are adapting to the realities of GenAI risk.

Key findings: 15% of companies faced a GenAI-related security incident this year; only 4% report high confidence in their defenses; and talent shortages remain the top barrier to preparedness.

The message is clear: generative AI adoption is accelerating faster than the guardrails to secure it.

👉 Read the full Lakera 2025 GenAI Security Readiness Report

From new browser threats to infrastructure breaches, and a sobering look at enterprise readiness, this week’s stories reveal both the promise and fragility of AI’s next chapter.

See you next week!

[the_rock]

All great points Val.