This week, security researchers confirmed something the industry had been bracing for: a single attacker used hundreds of autonomous AI agents to breach at least 395 organizations across 48 countries — compromising 11 of them in just 26 seconds at peak speed. Separately, a security startup disclosed sandbox-escape flaws in the very AI coding assistants — Claude Code, OpenAI Codex, and Cursor — that automation builders and no-code developers increasingly rely on every day.
Neither story is about some far-off, theoretical risk. Both are about tools already sitting in production stacks. If your team uses AI agents to write code, manage workflows, or touch customer data, here's what actually happened, and what it should change about how you operate.
According to research from GreyNoise, reported by BleepingComputer, The Hacker News, and The Register, a Russian-speaking threat actor built a private lab to develop exploits for two vulnerabilities in PaperCut NG/MF print management software (CVE-2026-81578 and CVE-2026-82078). Rather than running the attack by hand, the operator wired up an army of AI agents on an OpenAI Codex-style harness paired with a DeepSeek model — reportedly chosen specifically because it lacked the safety guardrails that frontier US models impose on offensive security requests.
The results were staggering. The agents went from an empty workspace to full remote code execution against a real victim in under four hours. Two hours after that, they had domain administrator access. Once the campaign launched in earnest, the agents compromised 11 separate organizations in 26 seconds. By the time researchers caught up, at least 440 PaperCut instances tied to 395 distinct organizations — nearly half of them in education — had been touched, with credentials harvested from 280 victims and full admin control gained at 12.
This lines up with what Anthropic disclosed in its own September 2026 threat intelligence report: state-linked and criminal groups are increasingly using AI not just to assist attacks, but to orchestrate and execute them end-to-end, with a human doing little more than supervising.
The same week, stealth security startup Accomplish disclosed a class of "configuration-based sandbox escape" bugs affecting Claude Code, OpenAI Codex, Cursor, Gemini CLI, and Antigravity. In short: a maliciously crafted .git config file could trick these coding agents into executing attacker-controlled code outside their intended sandbox — turning a tool meant to help you ship faster into an entry point.
OpenAI and Cursor shipped fixes within about a week of being notified. Anthropic's patch took roughly 50 days and 30 releases. The lesson isn't that any one vendor is careless — it's that the industry's guardrails around agentic coding tools are still catching up to how fast those tools are being adopted.
If your stack includes AI coding assistants, agentic workflow builders, or any tool that can act autonomously on your behalf, this isn't abstract:
AI agents move at machine speed, not human speed. A misconfiguration or unpatched CVE that used to give you days to respond can now be fully exploited in hours. Patch cadence has to tighten accordingly.
Agentic coding tools need the same scrutiny as any third-party dependency. Sandbox escapes in Claude Code, Codex, and Cursor mean these tools should run with least-privilege access, isolated environments, and no direct line to production credentials or client data.
"AI-assisted" attacks are now cheaper than ever to run. Open-weight models without safety restrictions lower the bar for attackers who previously needed real skill. Assume your defenses will be probed by automation, not just people.
RAG assistants on your documents, AI agents that act in your tools, and LLM features inside your product.
Our cutting-edge features simplify collaboration and creativity, making your workflow intuitive and efficient. Transform your vision into reality effortlessly with Hadidiz Flow.
