Anthropic Launch Claude Code Security

On February 20, 2026, Anthropic unveiled Claude Code Security, a tool that fundamentally changes the approach to finding security vulnerabilities in software.

Unlike traditional static analyzers that work by comparing code against databases of known vulnerability patterns, this tool leverages the advanced AI model Claude Opus 4.6 to analyze code in a manner closely resembling a human security analyst. It can trace data flows across an entire application, understand the relationships between components, and uncover sophisticated flaws in business logic or access control that rule-based tools would simply miss.

The tool is currently available as a limited research preview for Enterprise and Team tier customers.

Behind the project is more than a year of work by the internal Frontier Red Team, a group of approximately 15 researchers specializing in testing the boundaries of Anthropic’s most advanced AI systems.

During internal testing, Claude Opus 4.6 identified over 500 critical security flaws in production open-source projects. Many of these vulnerabilities had remained undetected for decades despite repeated expert reviews. Anthropic is currently undergoing a responsible disclosure process in collaboration with the maintainers of the affected projects.

One of the key aspects of Claude Code Security is its emphasis on minimizing false positives — a weakness that plagues many existing security tools. Each identified finding undergoes a multi-stage verification process in which the AI model attempts to disprove its own findings, filtering out inaccurate results before they reach the analyst. Validated findings are rated by severity level and detection confidence. All results are then displayed in a clear dashboard where the security team can examine the specific issue, the proposed fix, and the vulnerability context.

Anthropic emphasizes the human-in-the-loop principle: no fix is applied automatically; developer approval is always required. This approach reflects the reality that even the best AI can sometimes misinterpret code context.

The launch of Claude Code Security is a direct response to the growing threat posed by attackers’ use of AI. State-sponsored hacking groups and criminal organizations are already actively using generative AI in their operations, particularly in the preparatory phases of attacks — from reconnaissance to automated vulnerability scanning. Anthropic is therefore providing a tool intended to give defenders the upper hand.

The release of Claude Code Security triggered an immediate reaction in financial markets. Shares of leading cybersecurity firms saw significant declines. CrowdStrike dropped nearly 8%, Okta fell more than 9%, and Cloudflare also declined approximately 8%. Investors fear that AI agents capable of autonomously finding and fixing vulnerabilities may gradually replace traditional security products focused on threat detection.

For penetration testers, red teamers, and security consultants, this development presents both a challenge and an opportunity. AI tools like Claude Code Security will undoubtedly change the dynamics of the industry. Routine vulnerability scanning will become faster and more accessible, which will increase the pressure on the quality and depth of manual security audits.

DJI ROMO Robot Vacuum: A Gateway Into Thousands of Homes

Sammy Azdoufal, Director of AI Strategy at Emerald Stay, simply wanted to control his new DJI ROMO robot vacuum using a PlayStation 5 gamepad. With the help of the AI assistant Claude Code, he reverse-engineered the communication protocols of the DJI mobile app, built his own control application, and connected to the company’s servers. Instead of one vacuum, approximately 7,000 devices from 24 countries responded. Azdoufal was then able to remotely control other people’s vacuums, view live camera feeds, listen through their microphones, and generate detailed 2D floor plans of apartments and houses. All he needed was a mere 14-digit serial number.

At the core of the entire incident is the MQTT protocol, which DJI uses for communication between ROMO devices and its cloud platform DJI Home. This widely used IoT protocol operates on a publish-and-subscribe model through a central broker. The security of the entire system depends on whether the broker correctly enforces which clients may subscribe to which specific topics.

In DJI’s case, however, there were no topic-level access control lists (ACLs). Azdoufal extracted the authentication token from his own vacuum, connected to the MQTT broker, and the server granted him unrestricted access to data from all devices on the network. In just nine minutes, his scanner captured over 100,000 messages containing serial numbers, cleaning routes, battery statuses, and even camera recordings. DJI claimed the communication was encrypted using TLS, which is true — but TLS only protects data in transit. Once an attacker authenticates to a broker with no access controls, transport-layer encryption is entirely irrelevant because everything is readable at the application layer.

The company acknowledged it was aware of the backend authorization validation issue and announced that a fix was deployed in two automatic updates on February 8 and 10, 2026.

Azdoufal pointed out that some vulnerabilities remain unpatched even after the updates, including the ability to view the camera feed of one’s own ROMO without entering the security PIN. He declined to publicly describe another, more severe vulnerability until DJI patches it. The company promised a fix within weeks — which is quite a long time for a device equipped with a camera and microphone that autonomously maps the interiors of people’s homes.

Vulnerabilities in Password Managers

A team of cryptographers from ETH Zurich and Università della Svizzera italiana in Lugano conducted a thorough security analysis of three popular cloud-based password managers: Bitwarden, LastPass, and Dashlane. The results of the research, to be presented at the prestigious USENIX Security 2026 conference, revealed a total of 25 different attack scenarios. The researchers identified 12 attacks on Bitwarden, 7 on LastPass, and 6 on Dashlane. The most severe of these attacks allowed reading and modifying the entire vault contents — a complete compromise of stored passwords, payment details, and other sensitive data.

The research methodology involved simulating a compromised server. The researchers set up their own servers that behaved like compromised password manager infrastructure and observed what happened during normal user interactions.

A key finding was that no special techniques were needed for successful attacks. Ordinary operations that users or their browsers perform every day were sufficient. The weaknesses were found primarily in account recovery mechanisms, vault sharing features, and support for legacy cryptographic methods maintained for backward compatibility.

Professor Kenneth Paterson from ETH Zurich expressed surprise at the severity of the vulnerabilities found. According to him, the team had previously discovered similar issues in other cloud services, but expected a significantly higher level of security from password managers.

Interestingly, 1Password, which was not the primary subject of the study, stated that the identified attack vectors are already documented in its publicly available security design.

Firewalls as the Primary Ransomware Entry Point

Barracuda Networks published its annual Managed XDR Global Threat Report. According to an analysis of more than two trillion IT events collected throughout 2025, firewalls were exploited in 90% of ransomware attacks — either through unpatched software or compromised accounts with privileged access.

One of the report’s critical findings is the speed at which modern ransomware attacks unfold. The fastest recorded case involved the Akira ransomware, which managed to reach data encryption in just three hours. Attackers needed only ten minutes to move laterally to other systems after compromising a firewall. The data further shows that 96% of incidents involving lateral movement ended with ransomware deployment.

The report also highlights the persistent risk of unpatched software. The most frequently detected vulnerability was CVE-2013–2566, a flaw in the outdated RC4 encryption algorithm that has existed for thirteen years and still appears in legacy systems, old servers, and embedded devices. This single vulnerability served as the entry point for more than 320 attacks. Overall, one in ten detected vulnerabilities had a known exploit. Attackers are also actively using legitimate IT tools to mask their malicious activities.

Another significant trend identified in the report is a dramatic increase in supply chain and third-party attacks. In 2025, the share of incidents related to the supply chain rose to 66%, up from 45% the previous year. Attackers are deliberately seeking weaknesses in third-party software to bypass the defensive mechanisms of target organizations.