Daily Cybersecurity News - July 9, 2026
AI Agents Tricked Into Running Malicious Code
AI coding agents meant to find bugs can instead run attacker code on your machine.
The Friendly Fire proof of concept targets Anthropic's Claude Code. Attackers hide malicious instructions inside open source repositories that the agent scans.
The Friendly Fire PoC works against Claude Code and OpenAI Codex when their opt-in autonomous modes are enabled.
The demo uses a README instruction that convinces the agent to run an apparently legitimate script, which then launches a hidden payload..
Hackers exploit Roundcube flaw to spy on academic researchers
A China-linked cluster is hitting university email servers to steal researcher credentials and drop backdoors.
They exploit a Roundcube vulnerability to access accounts at US and Canadian schools. The attackers then install custom malware for long-term access.
The play reuses familiar webmail flaws, but focuses tightly on academic targets instead of typical government or corporate ones.
Multiple universities were breached over several months with credential theft as the primary goal.
GodDamn Ransomware Uses BYOVD to Smite US Companies
GodDamn ransomware is hitting US companies and using a Microsoft-signed kernel driver to disable security tools on the fly.
They bring their own vulnerable driver, load it legitimately, then leverage it to terminate AV processes before encryption starts. Victims are primarily US-based organizations across multiple sectors.
The PoisonX driver is unusual because it appears to be malicious code that nevertheless received a legitimate Microsoft Hardware Compatibility signature.
Symantec observed it in ransomware activity, but the report does not establish deployment at scale or say Microsoft revoked its signature..
Lone Attacker Uses AI to Breach AWS Cloud Environment
A lone attacker chained AI workflows with stolen credentials to breach a large Amazon customer's AWS environment in under 72 hours.
The intruder exploited misconfigured AI pipelines and cloud weaknesses to move laterally and exfiltrate data for extortion.
The approach mixes fresh AI tooling with classic credential theft and cloud misconfigurations rather than novel zero-days.
Single-person operation lasted three days before the victim noticed.
AI Coding Tools Tricked Into Hacking Developer Machine
A new attack tricks AI coding assistants into running malicious code on developer machines.
Wiz demonstrated the technique against multiple popular AI coding assistants in controlled tests.
Malicious AI Agent Skills Slip Past Scanners
AI coding agents pull add-on skills from public marketplaces just like npm packages, and some of those skills carry malicious instructions.
Researchers tested evasion techniques against 1,613 malicious agent skills and multiple install-time scanners.
SkillCloak’s packing technique bypassed every tested scanner more than 90% of the time, while its rewriting technique bypassed most static scanners more than 80% of the time.
The scanners focus on code and ignore plain-English instructions that tell the agent to exfiltrate data or run unauthorized commands.
Researchers posted the malicious skills on GitHub for defenders to test against.