IT News Roundup: Ghostcommit AI Attack, SAP Critical Flaws, and On-Device AI Breakthroughs - July 15, 2026

The latest round of IT news covers a shifting threat landscape in AI-assisted development, critical enterprise software vulnerabilities, and major advances in edge AI deployment. A new supply-chain attack technique targets AI code review tools, SAP patches a maximum-severity flaw affecting thousands of enterprise deployments, and Apple explores breakthrough model compression for on-device intelligence.

Ghostcommit Attack Hides Prompt Injection in Images to Fool AI Code Reviewers

Researchers from the ASSET Research Group at the University of Missouri-Kansas City disclosed a novel supply-chain attack technique dubbed Ghostcommit that conceals malicious prompt-injection instructions inside PNG image files. The attack exploits a blind spot in AI-powered code review tools: many automated reviewers, including CodeRabbit and Bugbot, skip image files entirely during their analysis.

In the proof-of-concept, a seemingly benign PNG file is included in a pull request alongside an AGENTS.md configuration file that references the image. When an AI coding agent processes the repository, the hidden instructions in the image trick it into reading sensitive files like .env and exfiltrating secrets — encoded as a list of numbers to avoid detection. The technique bypassed multiple AI code review systems without raising any flags.

The finding underscores a growing class of threats targeting the AI-assisted development workflow. As organizations increasingly rely on automated code review and AI coding agents, attackers are adapting their tactics to exploit gaps in how these tools process non-text files. Developers are advised to audit which tools handle image files in their CI/CD pipelines and to restrict AI agent access to sensitive configuration files.

Source: BleepingComputer | Malwarebytes

SAP July Patch Day Addresses Critical CVSS 9.9 Memory Corruption in NetWeaver

SAP released its July 2026 security updates, patching 16 vulnerabilities across multiple enterprise products. The most severe, tracked as CVE-2026-44747 with a maximum CVSS score of 9.9, is a memory corruption vulnerability in the SAP NetWeaver Application Server ABAP — a component deployed in thousands of enterprise environments worldwide.

The update also addresses two additional critical flaws: one in SAP Commerce Cloud and another in SAP AppRouter. Memory corruption bugs in enterprise application servers can potentially enable remote code execution, making timely patching essential. SAP recommends that all customers apply the July 2026 security updates as soon as possible, particularly for systems exposed to untrusted networks.

Organizations running SAP NetWeaver, Commerce Cloud, or AppRouter should verify their patch status immediately. The breadth of this month's update — spanning multiple products with three critical-severity entries — reflects the increasing pressure on enterprise software vendors to address complex attack surfaces.

Source: SecurityWeek | BleepingComputer

Apple in Talks with PrismML to Run 27-Billion-Parameter AI Models on iPhone

Apple is reportedly in discussions with PrismML, a startup specializing in AI model compression, to enable much larger language models to run entirely on-device. PrismML has demonstrated the ability to shrink Alibaba's Qwen 3.6 model — which has 27 billion parameters — to fit and operate on an iPhone 17 Pro, significantly exceeding the capacity of Apple's current on-device Apple Foundation Models.

The compressed model reportedly handles complex conversations, advanced reasoning tasks, and autonomous agent workflows while running locally without cloud dependency. PrismML plans to open-source two compressed versions of the model on July 14, designed for everyday devices including iPhones, MacBooks, and Nvidia-powered PCs. The startup's longer-term goal is to compress trillion-parameter models for edge deployment.

This development signals a shift in the on-device AI race. Running larger models locally improves privacy, reduces latency, and eliminates cloud inference costs — but it demands breakthroughs in model compression, quantization, and hardware optimization. For homelab enthusiasts, the techniques behind PrismML's compression could have implications for running capable models on consumer-grade hardware.

Source: CNBC | MacRumors

China Reverses Policy, Allows Top AI Firms to Purchase Nvidia H200 Chips

After months of restricting Chinese companies from buying Nvidia's H200 AI accelerators — even after the US government approved limited exports — Beijing has signaled a policy reversal. Reports indicate that China plans to allow leading domestic AI firms including Alibaba, ByteDance, and DeepSeek to purchase a limited supply of the chips.

The shift comes as Chinese AI companies face a severe computing capacity shortfall that domestic hardware alternatives have not been able to fully resolve. Earlier this year, the US Department of Commerce's Bureau of Industry and Security revised its export licensing policy to review H200 applications on a case-by-case basis, with conditions limiting the proportion shipped to China relative to US customers.

This development reflects the ongoing tension between US export controls aimed at curbing advanced AI development in China and the practical realities of global semiconductor supply chains. IT infrastructure planners should monitor how this policy shift affects GPU availability and pricing in the broader market.

Source: Android Headlines | South China Morning Post

Discourse Platform Patches SVG-Based XSS Vulnerability (CVE-2026-53962)

The open-source discussion platform Discourse has patched a cross-site scripting vulnerability stemming from insufficient SVG sanitization in file upload and user avatar handling. Affected versions include those prior to 2026.6.0, 2026.5.1, 2026.4.2, and 2026.1.5.

An attacker could exploit the flaw by uploading a crafted SVG file that executes arbitrary JavaScript in the browsers of users who visit specific URLs. Community forum administrators running self-hosted Discourse instances should upgrade to the patched versions as soon as possible. The vulnerability is particularly relevant for homelab operators who run Discourse as a community platform.

Source: NVD

Open Source Updates: Rust 1.97 Ships with New Symbol Mangling, Wireshark Patches 12 Security Issues

Two notable open source releases landed this week. Rust 1.97 introduced a new default symbol mangling scheme — a change that, while subtle, has significant implications for library ABI stability and cross-crate debugging in the Rust ecosystem. Existing Rust projects may need to review build configurations after upgrading.

Meanwhile, Wireshark 4.6.7 was released with fixes for 12 security vulnerabilities alongside updated protocol support. Network analysts and homelab operators who use Wireshark for packet inspection should apply the update promptly.

Source: Tech on Linux Stuff