Anthropic Widens Glasswing As Macos Threats Rise Arabian Post

(MENAFN- The Arabian Post) clearfix"> Anthropic is expanding access to Claude Mythos Preview through Project Glasswing, widening a controlled cybersecurity programme at a time when North Korean-linked hackers are sharpening attacks against macOS users in the financial, venture capital, Web3 and cryptocurrency sectors.

The San Francisco-based AI company plans to increase the number of Project Glasswing partners from about 50 to roughly 200 organisations across more than 15 countries, broadening access to a model designed to identify and help fix serious software vulnerabilities. The expansion comes as cyber researchers track a multi-stage macOS intrusion campaign attributed to Sapphire Sleet, a North Korean state-sponsored group also known as BlueNoroff and UNC1069.

The two developments underline a faster-moving security contest: defenders are turning to more powerful AI systems to find flaws before attackers exploit them, while state-linked cyber groups are using social engineering, fake software updates and trusted system tools to break into high-value targets without relying on conventional software exploits.

Project Glasswing was launched after Anthropic concluded that Claude Mythos Preview showed unusually strong capabilities in cybersecurity work, including vulnerability discovery, exploit analysis and testing of foundational systems. The company has said it does not plan to make the preview model generally available, citing the need for safeguards that can block dangerous outputs while still allowing legitimate security work.

Partners in the programme are expected to use the model for vulnerability detection, black-box testing of binaries, endpoint security and penetration testing. The effort includes model usage credits worth $100 million and funding support for open-source security groups, including Alpha-Omega, OpenSSF and the Apache Software Foundation. Anthropic has also said it will report publicly on lessons from the programme, including vulnerability remediation where disclosure rules allow.

See also AI bug hunt strains patch pipelines

The expansion is drawing attention from financial institutions, government agencies and critical infrastructure operators because advanced AI systems could change the pace at which software weaknesses are found and weaponised. Project Glasswing partners have already identified more than 10,000 flaws rated highly or critically severe, intensifying debate over whether restricted access to such models can strengthen defence without creating new risks.

That debate is gaining urgency as Sapphire Sleet continues to target people and organisations linked to digital assets. The group has been active since at least 2020 and is primarily associated with financially motivated operations aimed at cryptocurrency theft, blockchain-related intellectual property and access to high-value systems.

The macOS campaign uses professional lures rather than technical exploitation at the entry stage. Targets are contacted through platforms such as LinkedIn, Telegram, email or other business channels, where attackers pose as recruiters, investors or potential partners. Victims are then pushed towards a supposed video meeting and instructed to run a fake Zoom SDK update.

The malicious file, named to resemble a legitimate Zoom update, opens in macOS Script Editor. Its visible content appears benign, but the harmful code is hidden below thousands of blank lines. Once executed, the script uses trusted macOS tools such as osascript and curl to retrieve additional payloads, helping the attackers move through the infection chain while reducing the likelihood that the victim will notice anything unusual.

Security teams have identified credential theft, privacy control abuse, persistence and data exfiltration as core stages of the operation. One fake application presents a native-looking password prompt to capture the user's login credentials. The malware then abuses macOS privacy controls, including the Transparency, Consent and Control database, to expand access and suppress prompts that might otherwise alert the user.

See also AntV package breach widens npm alert

The campaign is designed to collect cryptocurrency wallets, browser extension data, Telegram session information, SSH keys, Apple Notes content and other sensitive files. Some malware components stage stolen material in temporary directories before compressing it and uploading it to remote servers. Other components create persistence through launch daemons, allowing malicious code to run when the system starts.

A separate investigation into UNC1069 activity found the group using fake Zoom meetings, compromised Telegram accounts and ClickFix-style prompts that instruct victims to run troubleshooting commands. The same playbook has also shown signs of AI-enabled social engineering, including the possible use of deepfake-style video in cryptocurrency-sector lures.

Apple has implemented protections to help detect and block infrastructure and malware associated with the macOS campaign after receiving technical details from security researchers. The case shows that macOS protections such as Gatekeeper, notarisation checks and quarantine enforcement can be weakened when users are persuaded to execute scripts or terminal commands themselves.

For financial and crypto organisations, the risk is especially acute because employees, developers, founders and investors often move between messaging apps, video calls, code repositories and wallet infrastructure. A single compromised endpoint can expose credentials, private keys, session tokens and internal project data.

Project Glasswing's expansion reflects the defensive side of the same shift. AI models that can find complex software flaws at scale may help organisations harden systems faster, but their capabilities also raise questions about misuse, access controls and disclosure rules. As North Korean-linked groups continue to refine macOS-focused intrusion methods, the race between AI-assisted defence and human-led deception is becoming a central test for the cybersecurity industry.

MENAFN03062026000152002308ID1111203953