Prompt Injection Cracks Claude Code as Gitea and Next.js Bugs Get Exploited

AI & Machine Learning

Security researcher Johann Rehberger published a working indirect prompt-injection attack against Claude Code’s Auto Mode, which Anthropic recently made the default for Pro, Max and Team users. In his demonstration, simply asking the agent to summarize a web page can lead it to download and unpack a zip archive and then execute attacker-controlled code, because a benign-looking import base64 step quietly pulls in a malicious struct.py planted from the archive. Rehberger reports success rates between 60% and 80%, and in one variant the payload launches a second headless agent via claude -p with its own tools and context. Anthropic’s response is that Auto Mode is a best-effort convenience classifier rather than a security boundary, and that OS isolation and network egress controls remain the real containment layer. The episode underlines how quickly agentic coding tools expand the attack surface as they gain autonomy. Source: The Register Verified: True

New analytics data quantified how AI-driven search is reshaping web traffic. A study of 54 advertisers found organic search sessions down 10.5% as Google’s AI Overviews became common, with 46 of the 54 seeing declines, while referral traffic from AI assistants such as ChatGPT, Copilot, Gemini and Perplexity rose 163%. The same period saw ChatGPT referrals to publishers fall roughly 52% between July and August, coinciding with a 53% jump in AI citations to Wikipedia, Reddit and TechRadar. Despite lower volumes, visitors arriving from AI tools completed valuable on-site actions at about 1.5 times the organic-search rate, and AI-attributed key events climbed 335%. The numbers reinforce a structural shift: fewer clicks overall, but higher intent among the users who do arrive. Source: MarketingProfs Verified: True

Consumer Hardware

Adobe launched a beta “AI Assisted Editor” mode in Photoshop, a separate workspace built around a natural-language prompt bar rather than the traditional toolset. It consolidates generative features—Prompt to Edit, Remove Background, Generative Expand, Generative Fill, Generative Upscale and an AI Markup tool that turns a rough doodle into an edit—into a single simplified panel aimed at less experienced users. The mode is reached from a new toggle on Photoshop’s menu bar and can be hidden from settings. Adobe also shipped “Instruct Edit with Masks,” powered by Firefly Image 5, and a new Light Adjustment Layer that brings exposure and contrast controls into the layers workflow. The move continues Adobe’s push to make prompt-first editing a mainstream on-ramp without abandoning its professional base. Source: 9to5Mac Verified: True

Google expanded AI Mode in Search with a set of travel-planning tools. Users can now describe a trip and get flight options with live pricing from more than 300 airlines and booking sites, ask AI Mode to track specific fares and email them when prices drop, and discover and book hotels in a conversation, with inventory from partners including Booking.com, Expedia, Hilton, Marriott, IHG and Priceline. A points-and-miles feature shows what a flight or hotel would cost in loyalty currency. Hotel booking is rolling out first in the US in English. The features deepen Google’s effort to turn AI Mode into an end-to-end assistant and raise the competitive stakes with dedicated travel apps and OpenAI’s shopping ambitions. Source: TechCrunch Verified: True

Cybersecurity

A critical remote-code-execution flaw in the self-hosted Git service Gitea, CVE-2026-60004, moved from disclosure to active exploitation. The bug, rated CVSS 9.8, lives in Gitea’s diff/patch handling and lets an attacker write custom Git hooks into a repository that then run arbitrary shell commands as the Gitea process user; on default installs with open registration enabled, that can be reached by a remote, effectively unauthenticated attacker. CISA added it to the Known Exploited Vulnerabilities catalog on 25 August, and researchers observed in-the-wild attacks dropping miner-like payloads. Gitea patched the issue in version 1.27.1 in late July, but more than 8,300 internet-exposed instances remained unpatched. Administrators are urged to update immediately and disable open self-registration where it is not needed. Source: The Hacker News Verified: True

Vercel pulled forward an emergency Next.js security release to 25 August, shipping versions 16.3.3 and 15.5.24 to fix two flaws that each allow unauthenticated remote code execution. One originates in the upstream libheif dependency and can be triggered when Image Optimization processes a crafted AVIF file; the patched releases disable AVIF optimization entirely until the upstream fix lands. The second, CVE-2026-75604, is a path-traversal issue affecting Next.js servers running on Windows filesystems that use the Pages or App Router without Cache Components. Applications hosted on Vercel were already mitigated and need no action, but self-hosted deployments—especially on Windows—should upgrade urgently given Next.js’s tens of millions of weekly downloads. Source: The Hacker News Verified: True

Enterprise Infrastructure

Marvell Technology reported fiscal second-quarter results on 27 August, posting record revenue and raising its outlook for both fiscal 2027 and 2028 on the strength of custom AI silicon bookings—the second straight quarter it has lifted guidance. Data-center revenue rose 46% year over year and now accounts for about 79% of the company’s business, as hyperscalers accelerate efforts to design their own accelerators rather than depend solely on Nvidia. Management guided to roughly $3.15 billion in revenue for the current quarter. The results follow Marvell’s recently disclosed $12.2 billion custom-chip agreement with Google and add to evidence that the merchant-silicon and custom-ASIC layer of the AI buildout is still expanding rapidly. Source: Quartz Verified: True

Policy & Regulation

More than 80 British performers launched a campaign, “Save Our Voices Now,” pressing the UK government to make a person’s voice a protected statutory right, comparable to existing protections for name and likeness. Backers including Nicola Coughlan, Matt Lucas and Hugh Bonneville signed an open letter to the prime minister and launched a public petition, arguing that modern voice-cloning systems need only about three seconds of audio to convincingly imitate a performer. Audiobook narrator Alice Sockett, who co-founded the campaign, called unauthorized cloning “an existential threat” to the narration industry. Organizers point to recent Danish reforms granting individuals ownership of their likeness, including voice, as a model. The campaign adds pressure as the UK weighs how copyright and personality rights should adapt to generative AI. Source: The Express Tribune Verified: True