AI security

Pluto Security counts 31 malicious Hugging Face repositories in two waves that abuse trust_remote_code to run a miner; the second-wave flagship model scanned clean and reached 28,412 downloads.

Kernel docs, GNOME, Google, Go and PHP each handled AI-found vulnerabilities differently in October, from report volume to CVE assignment and embargo timing.

A Concordia University benchmark found Cursor, Claude Code, and Codex Desktop let malicious GitHub issues bypass their safety guardrails in up to 79.2% of attempts, with GPT-based agents far more vulnerable than Claude Code.

A new paper testing seven defenses against AI-hallucinated software packages finds only RAG and Self-Refine hold up under adversarial prompting; Ruby stays the most vulnerable language.

A crafted link could seed attacker instructions into a user's Rovo session, letting the assistant's browsing agent exfiltrate Jira and Confluence data.

Anthropic's new beta feature routes every Claude Enterprise prompt through a customer-run security server for an allow-or-deny verdict before inference runs.

A coalition led by Iowa's Brenna Bird warns OpenAI it may have broken consumer-protection and data-privacy laws after an AI agent escaped a test sandbox and hacked Hugging Face.

A Singapore researcher disclosed CVE-2026-53264, an AI-assisted Linux kernel use-after-free that escalates local access to root on CentOS Stream 9.

The Linux Foundation and roughly 20 companies including AWS, Anthropic, Google, IBM and major banks launched Akrites, a coordinated incident-response alliance for open source software, after finding fewer than 5% of recently surfaced vulnerabilities have been patched.

Hugging Face disclosed that an autonomous AI agent system breached its data-processing pipeline over a weekend, harvesting credentials before defenders turned to a self-hosted open-weight model to analyze the attack.

OpenAI says its internal attacker model GPT-Red cut GPT-5.6 Sol's direct prompt-injection failure rate sixfold and uncovered a new 'fake chain of thought' attack class.

Sysdig says JadePuffer used an LLM agent to exploit Langflow flaw CVE-2025-3248, run 600-plus payloads, and encrypt 1,342 Nacos config items.