Remote Working Cyber Risks

Security researchers discovered two critical vulnerabilities in Anthropic's agentic AI coding tool, Claude Code. The vulnerabilities, tracked as CVE-2025-59536 and CVE-2026-21852, allowed attackers to achieve remote code execution and to compromise a victim's API credentials. The vulnerabilities exploit maliciously crafted repository configurations to circumvent control mechanisms. It should be noted that Anthropic worked closely with the security researchers throughout the process, and the bugs were patched before the research was published. 

What’s Notable and Unique 

• The configuration files .claude/settings.json and .mcp.json were repurposed to execute malicious commands. Because the configurations could be applied immediately upon starting Claude Code, the commands ran before the user could deny permissions via a dialogue prompt, or they bypassed the authentication prompt altogether. 

  
  

• .claude/settings.json also defines the endpoint for all Claude Code API communications. By replacing the default localhost URL with a URL they own, an attacker could redirect traffic to infrastructure they control. Critically, the authentication traffic generated upon starting Claude Code included the user's full Anthropic API key in plain text and was sent before the user could interact with the trust dialogue. 

  
  

• Restrictive permissions on sensitive files could be bypassed by simply prompting Claude Code to create a copy of the file's contents, which did not inherit the original file's permissions. A threat actor using a stolen API key could gain complete read and write access to all files within a workspace. 

Analyst Comments 

The vulnerabilities and attack paths detailed in the research illustrate the double-edged nature of AI tools. The speed, scale, and convenience characteristics that make AI tools attractive to developer teams also benefit threat actors who use them for nefarious purposes. Defenders should expect adversaries to continue seeking ways to exploit configurations and orchestration logic to increase the impact of their attacks. Organizations planning to implement AI development tools should prioritize AI supply-chain hygiene and CI/CD hardening practices. 

Sources 

• Caught in the Hook: RCE and API Token Exfiltration Through Claude Code Project Files | CVE-2025-59536 | CVE-2026-21852

After a slight lull in January, Akira and Qilin returned to dominating ransomware activity in February, collectively accounting for almost half of all engagements that month. The rest of the threat landscape remained relatively diverse, with a mix of persistent threats like INC and PLAY, older groups like Cl0p and LockBit, and newer groups like BravoX and Payouts King. Given current trends, the first quarter of 2026 will likely remain relatively predictable, with the top groups from the second half of 2025 continuing to operate at fairly consistent levels month to month.

Figure 1. Activity from the top 5 threat groups in February 2026

Throughout the month of February, analysts at Arete identified several trends behind the threat actors perpetrating cybercrime activities: 

• In February, Arete observed Qilin actively targeting WatchGuard Firebox devices, especially those vulnerable to CVE-2025-14733, to gain initial access to victim environments. CVE-2025-14733 is a critical vulnerability in WatchGuard Fireware OS that allows a remote, unauthenticated threat actor to execute arbitrary code. In addition to upgrading WatchGuard devices to the latest Firebox OS version, which patches the bug, administrators are urged to rotate all shared secrets on affected devices that may have been compromised and may be used in future campaigns.

  
  

• Reports from February suggest that threat actors are increasingly exploring AI-enabled tools and services to scale malicious activities, demonstrating how generative AI is being integrated into both espionage and financially motivated threat operations. The Google Threat Intelligence Group indicated that state-backed threat actors are leveraging Google’s Gemini AI as a force multiplier to support all stages of the cyberattack lifecycle, from reconnaissance to post-compromise operations. Separate reporting from Amazon Threat Intelligence identified a threat actor leveraging commercially available generative AI services to conduct a large-scale campaign against FortiGate firewalls, gaining access through weak or reused credentials protected only by single-factor authentication.

  
  

• The Interlock ransomware group recently introduced a custom process-termination utility called “Hotta Killer,” designed to disable endpoint detection and response solutions during active intrusions. This tool exploits a zero-day vulnerability (CVE-2025-61155) in a gaming anti-cheat driver, marking a significant adaptation in the group’s operations against security tools like FortiEDR. Arete is actively monitoring this activity, which highlights the growing trend of Bring Your Own Vulnerable Driver (BYOVD) attacks, in which threat actors exploit legitimate, signed drivers to bypass and disable endpoint security controls.

Sources

• Arete Internal

Security researchers identified a sophisticated evolution of the ClickFix campaign that aims to compromise legitimate websites before delivering a five-stage malware chain, culminating in the deployment of MIMICRAT. MIMICRAT is a custom remote access trojan (RAT) written in the C/C++ programming language that offers various capabilities early in the attack lifecycle. The attack begins with victims visiting compromised websites, where JavaScript plugins load a fake Cloudflare verification that tricks users into executing a malicious PowerShell script, further displaying the prominence and effectiveness of ClickFix and its user interaction techniques. 

Not Your Average RAT 

MIMICRAT displays above-average defense evasion and sophistication, including: 

• A five-stage PowerShell sequence beginning with Event Tracing for Windows and Anti-Malware Scan Interface bypasses, which are commonly used in red teaming for evading detection by EDR and AV toolsets.  

• The malware later uses a lightweight scripting language that is scripted into memory, allowing malicious actions without files that could easily be detected by an EDR tool.  

• MIMICRAT uses malleable Command and Control profiles, allowing for a constantly changing communication infrastructure. 

• The campaign uses legitimate compromised infrastructure, rather than attacker-owned tools, and is prepped to use 17 different languages, which increases global reach and defense evasion. 

Analyst Comments 

The ClickFix social engineering technique remains an effective means for threat actors to obtain compromised credentials and initial access to victim environments, enabling them to deploy first-stage malware. Coupled with the sophisticated MIMICRAT RAT, the effectiveness of this campaign could increase. Arete will continue monitoring for changes to the ClickFix techniques, the deployment of MIMICRAT in other campaigns, and other pertinent information relating to the ongoing campaign. 

Sources 

• MIMICRAT: ClickFix Campaign Delivers Custom RAT via Compromised Legitimate Websites

State-backed threat actors are leveraging Google’s Gemini AI as a force multiplier to support all stages of the cyberattack lifecycle, from reconnaissance to post-compromise operations. According to the Google Threat Intelligence Group (GTIG), threat actors linked to the People’s Republic of China (PRC), Iran, North Korea, and other unattributed groups have misused Gemini to accelerate target profiling, synthesize open-source intelligence, identify official email addresses, map organizational structures, generate tailored phishing lures, translate content, conduct vulnerability testing, support coding tasks, and troubleshoot malware development. Cybercriminals are increasingly exploring AI-enabled tools and services to scale malicious activities, including social engineering campaigns such as ClickFix, demonstrating how generative AI is being integrated into both espionage and financially motivated threat operations. 

What’s Notable and Unique 

• Threat actors are leveraging Gemini beyond basic reconnaissance, using it to generate polished, culturally nuanced phishing lures and sustain convincing multi-turn social engineering conversations that minimize traditional red flags.  

• In addition, threat actors rely on Gemini for vulnerability research, malware debugging, code generation, command-and-control development, and technical troubleshooting, with PRC groups emphasizing automation and vulnerability analysis, Iranian actors focusing on social engineering and malware development, and North Korean actors prioritizing high-fidelity target profiling. 

• Beyond direct operational support, adversaries have abused public generative AI platforms to host deceptive ClickFix instructions, tricking users into pasting malicious commands that deliver macOS variants of ATOMIC Stealer.  

• AI is also being integrated directly into malware development workflows, as seen with CoinBait’s AI-assisted phishing kit capabilities and HonestCue’s use of the Gemini API to dynamically generate and execute in-memory C# payloads.  

• Underground forums show strong demand for AI-powered offensive tools, with offerings like Xanthorox falsely marketed as custom AI but actually built on third-party commercial models integrated through open-source frameworks such as Crush, Hexstrike AI, LibreChat-AI, and Open WebUI, including Gemini. 

Analyst Comments 

The increasing misuse of generative AI platforms like Gemini highlights a rapidly evolving threat landscape in which state-backed and financially motivated actors leverage AI as a force multiplier for reconnaissance, phishing, malware development, and post-compromise operations. At the same time, large-scale model extraction attempts and API abuse demonstrate emerging risks to AI service integrity, intellectual property, and the broader AI-as-a-Service ecosystem. While these developments underscore the scalability and sophistication of AI-enabled threats, continued enforcement actions, strengthened safeguards, and proactive security testing by providers reflect ongoing efforts to mitigate abuse and adapt defenses in response to increasingly AI-driven adversaries. 

Sources 

• GTIG AI Threat Tracker: Distillation, Experimentation, and (Continued) Integration of AI for Adversarial Use