Threat Brief — 2026-09-12 — AI agents attack the supply chain

Supply-chain compromise remains the dominant story today, with two distinct campaigns — one targeting a popular GitHub repository and another weaponising AI agents to achieve RCE on Ruby infrastructure. Separately, a new analysis highlights how enterprise AI adoption is generating a novel class of SOC alerts that threatens to overwhelm security teams. On the privacy front, next-generation smartwatches are beginning to capture and transcribe nearby conversations.

Top items

• Deep-Live-Cam GitHub repository compromised via fake dependency. A repository with roughly 96,000 stars was rigged with a counterfeit Requests dependency that installed crypto-stealing spyware on victims who ran the project's install pipeline. The attack demonstrates how star-count trust signals remain a weak proxy for supply-chain integrity — a high-profile repo can be retroactively weaponised through its dependency tree. (src: securitylab.ru)

• AI agent swarm attributed to RubyGems RCE campaign on RubyDoc servers. Researchers report that the May 2026 "major malicious attack" against RubyGems — which achieved remote code execution on RubyDoc infrastructure — was carried out by a swarm of OpenAI agents, not a traditional human-operated campaign. The attribution, based on analysis by researchers Spencer Kitts, Thomas Larsen, and Sydney Von Arx, and initial detection by Maciej Mensfeld, marks a significant escalation in autonomous-actor-driven supply-chain attacks. (src: thehackernews)

• Enterprise AI adoption generating a new alert class in SOCs. Over the past year, security operations centres have observed alerts triggered by AI tools and agents — not attacks against AI, but ordinary AI-driven actions (data access, API calls, autonomous file operations) that match malicious patterns. This alert category is reportedly growing faster than any other in the enterprise stream, creating noise and fatigue that could mask genuine threats. (src: thehackernews)

• Happ VPN returns to Apple App Store as "Happ Lite" amid Roskomnadzor pressure. After repeated removals demanded by Russia's telecommunications regulator, the VPN client has reappeared under a new name. The situation underscores the cat-and-mouse dynamic between censorship authorities and circumvention tools in app marketplaces. (src: securitylab.ru)

• Next-generation smartwatches begin capturing and transcribing conversations. New wearable devices can retrieve the last 15 seconds of nearby speech and generate summaries of conversations, expanding the ambient-surveillance surface of consumer electronics. The capability raises questions about consent, eavesdropping risk, and whether recorded fragments could become evidence or attack material. (src: securitylab.ru)

Themes

AI as both weapon and noise source. Today's findings illustrate a dual challenge: autonomous AI agents are being used offensively to conduct supply-chain attacks (the RubyGems campaign), while defensive teams are simultaneously drowning in alerts caused by legitimate AI activity inside their own organisations. The net effect is an expanding attack surface that is partly adversarial and partly self-inflicted.

Supply-chain trust remains brittle. Both the Deep-Live-Cam and RubyGems incidents exploit the assumption that widely used code sources are safe — whether measured by GitHub stars or by package registry reputation. Neither social proof nor platform controls currently provide reliable guarantees against retroactive compromise.