Is PyTorch's 'Safe' Loading Setting Actually Safe?

Author

Taylor Hawes
August 31, 2026

Streamline identity threat detection and response

ADAudit Plus helps protect your identity infrastructure with capabilities like change auditing, logon tracking, privileged user monitoring, file auditing, attack surface analysis, user behavior analytics, response automation, backup and recovery, and compliance reporting across Active Directory, Entra ID, file systems, and more.

🔎 Cyber Watch 🔎

PyTorch Attack Chain Shows AI Security Controls Can Have Hidden Gaps

New research presented at Black Hat Asia 2026 has demonstrated a three-stage attack chain that can turn a malicious PyTorch model file into remote code execution. The researchers found that TorchScript files could reach a memory safety flaw even when PyTorch's weights_only=True setting was used. The attack was demonstrated against a live OpenSearch deployment, showing how weaknesses in AI model-loading workflows can create risks beyond traditional software supply chains. The story is another warning that AI frameworks, model files, and runtime environments need the same security scrutiny as application code.

Key takeaways

  • AI model files should be treated as potentially untrusted code.

  • Security settings can reduce risk without fully removing it.

  • AI supply chain security must cover frameworks and model-loading processes.

  • Organizations using PyTorch should review where and how third-party models are loaded.

🎙️ Tech Briefing On‑Air 🎙️

What Cybersecurity Teams Can Learn From the Racetrack

In this episode of Blackpoint Cyber's Return of the Mac, Kyle Kirkwood joins MacKenzie Brown and Oli Thordarson for a discussion on the links between motorsport and cybersecurity. The conversation focuses on preparation, rapid decision-making, data analysis, teamwork, and the role of specialists under pressure. Racing teams cannot wait until a problem becomes critical before responding, and the same principle applies to security operations. The episode offers a useful reminder that strong incident response depends on preparation and coordinated action as much as technical tools.

Takeaways

  • Fast decisions depend on preparation before an incident occurs.

  • Data is useful only when teams can turn it into timely action.

  • Clear roles and communication can reduce response delays.

  • Cybersecurity teams can benefit from studying how other high-pressure industries manage risk.

🤝 Partner Intel 🤝

ManageEngine ADAudit Plus Brings Identity Activity Into Focus

ManageEngine ADAudit Plus helps organizations track activity across Active Directory, Microsoft Entra ID, Windows servers, file servers, workstations, and NAS environments. The platform supports use cases including change auditing, privileged user monitoring, logon tracking, file auditing, compliance reporting, and threat detection. This broad visibility can help security teams investigate suspicious changes and identify risky activity across identity and Windows environments from a central platform. ManageEngine currently lists ADAudit Plus 8.7 Build 8711 as the latest version.

Why it matters

Identity-related changes are often buried among large volumes of routine activity. ADAudit Plus is designed to make changes, privileged actions, logons, and file activity easier to track, giving IT and security teams more context when investigating an incident or meeting compliance requirements.

🤖 AI Runtime 🤖

Tech Leaders Warn That AI-Powered Attacks Are Closing the Security Gap

More than 150 technology and cybersecurity companies have warned that the time available to strengthen defenses against AI-powered cyberattacks may be limited. A joint open letter calls for greater investment in cybersecurity and better defensive capabilities for critical infrastructure, including hospitals, water facilities, and other essential services. The companies argue that existing security practices will not be enough as AI systems become more capable of finding and exploiting weaknesses. Their message is clear: defenders must also gain access to stronger AI-powered tools and resources.

📊 By the Numbers 📊

Source: ScienceSoft

🗳️ Your Monday Take 🗳️

Cast your vote on our weekly poll.

What Would Hurt Your Security Team More for 24 Hours?

Login or Subscribe to participate in polls.

📩 We’ll share the results in the Friday issue.

Advertise with Comparitech
Does your business offer services or products in cybersecurity? Get your product seen by IT leaders and professionals.

Advertise with us →

Until Wednesday’s edition - Let’s keep that zero-day count at zero!