Anthropic stops AI spies, the new OWASP Top 10 and the rise of small-time ransomware

Anthropic’s Report on an AI-Powered Espionage Campaign

A recent report from Anthropic detailed the disruption of an espionage campaign where attackers used its model, Claude, to automate 80-90% of their operations. This included reconnaissance, writing exploit code, establishing backdoors, and exfiltrating data with minimal human input.

The reaction from the expert panel was that this development, while significant, was also “highly predictable.” The key innovation highlighted was not the model’s intelligence approaching AGI, but its advanced capability for tool orchestration and planning. The AI effectively acted as a sophisticated “script kiddie,” chaining together various open-source tools to execute the attack chain. The attackers socially engineered the model by adopting a “role play” pretext, posing as a security research firm to bypass some of its built-in guardrails.

From a defensive perspective, this represents the materialization of a threat incident response (IR) teams have been preparing for: “an adversary that really doesn’t sleep… and iterates at true machine speed.” While the AI can still make mistakes and hallucinate—such as generating non-working credentials—the primary threat comes from the scale, speed, and scope it enables. It significantly lowers the barrier to entry, empowering less-skilled actors to become highly effective.

The consensus is that defenders must “fight fire with fire.” Security teams are encouraged to adopt the same agentic AI techniques for proactive testing and defense. This is already happening, with defenders chaining defensive AI agents for tasks like log monitoring, memory forensics, isolating hosts, and even auto-analyzing regulatory frameworks to ensure compliance.

The New OWASP Top 10 for 2025

OWASP released its updated list of the top 10 most critical web application security risks, the first since 2021. Key changes include:

• Broken Access Control remains the #1 threat, indicating a disappointing lack of progress in a fundamental area.
• Security Misconfigurations rose from #5 to #2, highlighting a major attack vector that defenders have significant control over.
• Software Supply Chain Failures and Mishandling of Exceptional Conditions were added to the list, reflecting the modern, interconnected nature of applications and the importance of business logic flaws.

The panel views the list as a “reality check for the AI coding era,” serving as a predictive map of how attackers are successfully breaching modern environments. The key takeaway for defenders is to double down on the fundamentals. The rise of misconfigurations suggests a critical need for automation in system setup to reduce human error. Securing the development pipeline and gaining deep visibility into APIs and identity flows are paramount as AI-generated code becomes more prevalent.

Fragmentation in the Ransomware Ecosystem

The ransomware landscape has become increasingly decentralized. Successful law enforcement takedowns of large, established gangs have led to their dissolution, with former members scattering to form numerous smaller, more agile groups.

This fragmentation presents several new challenges:

1. A Dearth of Intelligence: It is much harder to attribute attacks and track the tactics, techniques, and procedures (TTPs) of many small, ephemeral groups compared to a few large, predictable ones.
2. Increased Volatility: Smaller groups have less of a “brand” to maintain, making them less reliable in negotiations and more likely to be chaotic or make mistakes, such as accidentally destroying a victim’s environment.
3. Law Enforcement Challenges: The traditional playbook for tracking large criminal enterprises with clear hierarchies and financial trails is ineffective. The new landscape is described as “whack-a-mole at a global scale.”

All panelists agreed they would rather face a few large, professionalized ransomware cartels than a multitude of small, unpredictable crews.

The Ongoing Cyber Insurance Debate

With cyber insurance payouts tripling in the UK in one year, the debate over its efficacy continues. The panel discussed whether insurance simply encourages attackers or if it helps strengthen defenses.

The consensus leaned toward insurance being a net positive, acting as a necessary safety net for business continuity and brand reputation. More importantly, insurers are becoming unofficial regulators of the cybersecurity industry. To reduce their own exposure, they are mandating stringent security controls as a prerequisite for coverage, such as:

• Multi-factor authentication (MFA)
• Immutable backups
• Evidence of incident response readiness and tabletop exercises

By doing so, the insurance industry is directly influencing and elevating the cybersecurity maturity of organizations, forcing them to adopt fundamental best practices to remain insurable. This provides a structured path to recovery and can prevent a security incident from becoming a business-ending event.

• Tags:
• Agentic AI
• Cybersecurity
• Incident Response
• Anthropic
• Claude
• Ransomware
• OWASP Top 10
• Cyber Insurance