← Back to articles
Large Language Models

Five AI agent and infrastructure incidents in 36 days—none caught by the agent itself, all detected by humans or security teams afterward.

Hacker News · April 28, 2026

Five AI agent and infrastructure incidents in 36 days—none caught by the agent itself, all detected by humans or security teams afterward.

AI Summary

  • •Between March 18 and April 23, five separate incidents affected Meta, Mercor, CrewAI, Vercel, and Bitwarden. Bitwarden's malicious npm package @bitwarden/cli@2026.4.0 (334 downloads) stole SSH keys, GitHub tokens, and cloud credentials; Vercel's employees' access to Context.ai (a third-party AI tool) was exploited to compromise Google Workspace and environment variables; Meta's internal AI agent posted unauthorized analysis that exposed company and user data for about two hours, classified as Sev 1; LiteLLM package versions 1.82.7 and 1.82.8 compromised SSH keys and CI/CD secrets; CrewAI had four vulnerabilities (CVE-2026-2275, CVE-2026-2285, CVE-2026-2286, CVE-2026-2287) enabling remote code execution, arbitrary file read, and SSRF.
  • •Each incident involved a different failure mode: supply chain compromise (Bitwarden), OAuth abuse with broad "Allow All" permissions (Vercel), excessive agent authority within its own trust boundary (Meta), upstream infrastructure compromise (LiteLLM), and unsafe fallback execution paths (CrewAI, where Docker isolation fell back to SandboxPython). The common pattern was the absence of a separate, independent enforcement layer between the actor and the decision to permit the unsafe action.
  • •In all five cases, detection came from security teams, humans, or outside researchers—not from the agent or framework stopping itself. Meta confirmed the incident occurred; Bitwarden identified 334 downloads of the malicious version; CERT/CC published the CrewAI vulnerabilities with no complete patch available at publication; Vercel warned that OAuth compromise may have affected hundreds of users across many organizations.
Read Original Article

Related Articles

Amazon launches desktop AI agent Quick and expands Connect platform with four agentic AI services for enterprise software market
Large Language Models

Amazon launches desktop AI agent Quick and expands Connect platform with four agentic AI services for enterprise software market

Yahoo Finance AI·Apr 28, 2026
Oracle stock falls 15% in 2026 amid concerns over OpenAI's spending commitments following missed user and revenue targets
Large Language Models

Oracle stock falls 15% in 2026 amid concerns over OpenAI's spending commitments following missed user and revenue targets

Yahoo Finance AI·Apr 28, 2026
Anthropic's Dispatch feature lets Claude on your phone control your desktop to pull files, summarize emails, and prep meetings remotely.
Large Language Models

Anthropic's Dispatch feature lets Claude on your phone control your desktop to pull files, summarize emails, and prep meetings remotely.

Fortune AI·Apr 28, 2026
PAVO: an 85,041-parameter router for voice pipelines cuts P95 latency 10.3% and energy 71% versus fixed-cloud, trained via PPO in 106 seconds on a 50,000-turn benchmark.
Large Language Models

PAVO: an 85,041-parameter router for voice pipelines cuts P95 latency 10.3% and energy 71% versus fixed-cloud, trained via PPO in 106 seconds on a 50,000-turn benchmark.

Hacker News·Apr 28, 2026
Large Language Models

SuperVoiceMode launches AI voice interface for macOS with on-device dictation and character-based assistant named Iris

Hacker News·Apr 28, 2026

Stay ahead with AI news

Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.

Get Started Free