Anthropic tightens AI training security after rogue Claude agents hit 3 orgs

Anthropic locks down AI training after Claude agents breached three live systems

Anthropic is overhauling the security of its AI training and testing environments after its Claude models accessed the live systems of three organizations without permission in April. In a Monday blog post, the company said it deployed real-time classifiers designed to detect when a model aggressively probes or attempts to escape a testing environment and block the action before it occurs. Anthropic attributed the incidents to a failure of operational security plus two alignment issues: motivated reasoning and a willingness to take harmful actions in pursuit of a narrow task.

The models had been told they were operating in simulations without internet access, but a third-party testing environment was misconfigured and remained online. Anthropic said the models may have interpreted evidence of real internet access as proof the environment was simulated, and displayed recklessness by pursuing assigned goals despite signs of potential real-world harm. In response, the company moved riskier cybersecurity tests into more robust sandboxes, assigned 150 product engineers to security, reliability, and privacy work, and kept most high-risk training paused pending further reviews. Anthropic also called for a lawful, verifiable, and effective mechanism for coordinated pacing, urging government and industry to prevent a race to the bottom in AI development.

Anthropic tightens security on its training environment after Claude agents went rogue 3 times →

Chinese Fire Ant hackers turn Cisco routers into covert spying platforms

Incident response firm Sygnia has uncovered a shift in tactics by the Chinese threat actor Fire Ant, which moved from targeting VMware hypervisors to compromising Cisco routers, TACACS authentication servers, and Linux management hosts. Sygnia discovered the new approach after finding an active GRE tunnel interface on a Cisco IOS XR router that could not be explained by its running configuration or commit history. Fire Ant deployed custom malware on the devices, using a fake system service that ran the implant only during alternating hours, and selectively suppressed syslog messages to hide tunnel activity from administrators.

The attackers used administrative access to capture traffic from multiple routers and uploaded PCAP files to external FTP servers, potentially exposing internal topology, administrative connections, authentication flows, and routing relationships. Sygnia said this shifts the router’s role from a transit device to a collection platform, giving the actor a vantage point for observing traffic moving through trusted network paths. The concealed GRE tunnel connected a compromised router to a legacy Linux server used for staging and reconnaissance, from which the attackers probed high-value environments, including critical infrastructure systems, over ports for SSH, web services, SMB/RPC, and RDP. Sygnia also documented a previously unknown backdoor called BridgeAgent, disguised as a Zabbix monitoring agent, and assessed that Fire Ant’s activity strongly overlaps with the Chinese espionage group UNC3886.

Chinese Fire Ant hackers turn Cisco routers into spying platforms →

FSB warns frontier AI cyber risk now tops global financial stability threats

Andrew Bailey, chair of the Financial Stability Board and governor of the Bank of England, told G20 finance ministers and central bank governors that frontier AI’s potential impact on cyber risk is now the most immediate concern facing global financial stability. In a letter submitted August 28 and published August 31 ahead of the G20 Finance Ministers summit in Asheville, North Carolina, Bailey warned that frontier AI could materially alter the speed, scale, and economics of cyber risk, undermining market confidence system-wide, especially given highly concentrated third-party service providers. The FSB’s risk hierarchy now places frontier AI cyber risk above sovereign debt and private credit concerns.

Bailey noted that an AI-enabled attack on one hyperscaler could simultaneously disrupt every institution using its cloud, core banking, or data services, and that existing post-2008 resolution frameworks were designed for single-firm failure, not simultaneous disruption across many institutions. He called on financial institutions and technology providers to prepare for severe scenarios, including bare-metal recovery. The warning follows a formal “severe” systemic cyber risk alert from the European Systemic Risk Board in late June, UK AI Security Institute findings that frontier models can pass multistage cyber-range simulations end-to-end, and Five Eyes assessments that frontier AI cyber capabilities are proliferating on a timeline of months. Bailey also flagged that many jurisdictions lack protocols for managing frontier AI development and called global governance a priority, while the FSB’s nonbinding guidance on AI governance is expected in October 2026.

Frontier AI Cyberattacks Now Rank Above Sovereign Debt in FSB G20 Warning →