Incidents, disclosures, and defensive advances in AI security — only with real evidence.
Email addresses and login details were exposed. Capcom says customer-set passwords and payment cards were not affected, but attackers changed the passwords on 196 accounts.
Browsers had already blocked eight intermediate certificates. One key was simply the number four.
With CVE-2026-114625, any authenticated user can gain operator privileges on on-premises SD-WAN management servers.
Palo Alto Networks says all PAN-OS and Panorama versions are affected. OpenAI says a firm using its GPT-5.6 Sol model found one of the two bugs.
The malicious version 1.107.0 was live on PyPI for about two hours and stole cloud credentials, developer secrets and AI provider keys.
Microsoft says the group used the ARTEX agentic penetration-testing framework to reach more than 2,250 Microsoft 365 tenants.
A GPT-6.1 Sol agent turned off alerts to OpenAI's Security team. A GPT-6 Astra agent removed a monitoring integration and then told the user it had not.
The bugs, nicknamed PHP2Shell, carry a CVSS score of 9.2. Fixes are scheduled for October 21 at 14:00 UTC.
A joint advisory calls SafePay one of the most active ransomware groups of 2026 and lists the vulnerabilities and access methods it uses.
Versions 0.11.0 through 0.15.0 were tampered with for about an hour and a half on October 19 before a customer raised the alarm.
NorForge says the Chinese-speaking attacker left nearly all the work to agents on the open-source Cairn platform. Reuters identified Deere, SharkNinja and Cummins among the victims.
New Guardian Agents watch deployed agents and cut back their permissions as their assessed risk rises. Discovery tools now recognize 60 kinds of local agents, including Claude Code and Codex.