CybersecurityVulnerability Management

Exploiting Vulnerabilities: Malicious ML Models on Hugging Face Use Flawed Pickle Format to Bypass Detection

Analyst 207|
Exploiting Vulnerabilities: Malicious ML Models on Hugging Face Use Flawed Pickle Format to Bypass Detection

Malicious ML Models Discovered on Hugging Face

Cybersecurity researchers have identified two malicious machine learning (ML) models hosted on Hugging Face, utilizing a unique method involving “broken” pickle files to evade detection. This discovery raises significant concerns regarding the security of machine learning models and the platforms that host them.

Details of the Discovery

According to a report by ReversingLabs researcher Karlo Zanki, the pickle files extracted from the PyTorch archives contained malicious Python content embedded at the beginning of the file. This technique allowed the models to bypass standard detection mechanisms.

Key Points

  • Two malicious ML models were found on Hugging Face.
  • The models used flawed pickle files to conceal their harmful content.
  • Malicious code was located at the start of the pickle files, facilitating evasion of detection.
  • This incident highlights vulnerabilities in the security of machine learning frameworks and repositories.

Conclusion

The discovery of these malicious models underscores the importance of vigilance in the cybersecurity landscape, particularly concerning the integrity of machine learning resources. Users and developers must remain cautious and implement robust security measures to protect against such threats.

Cyber SecurityMachine Learning
Related Intelligence

Continue Reading

Smartphone on a neutral surface with blank screen and hint of a hand nearby.

Biometric Age Verification Shifts to On-Device Processing

With data breaches skyrocketing - a record 3,322 in the US last year alone - and over 30 age assurance laws now in force worldwide, the need for robust biometric age verification has never been more pressing. As governments crack down on underage access to social media, on-device processing is emerging as a crucial solution.

Analyst 207
Laptop on a minimalist desk with a potted plant and stack of paper in soft natural light.

WordPress Discloses Core Flaw Enabling Unauthenticated Code Execution

WordPress has patched a critical flaw that allowed hackers to execute code remotely without authentication, releasing versions 6.9.5 and 7.0.2 to fix the vulnerability. The update addresses a REST API batch-route confusion and SQL injection issue that could be triggered by a simple HTTP request.

Analyst 207
Rows of computer servers and networking equipment in a dimly lit data center with one server highlighted.

OpenSSL Flaw Exposes Servers to Memory Exhaustion Attacks

A newly discovered OpenSSL flaw, dubbed HollowByte, leaves unpatched servers vulnerable to memory exhaustion attacks, where a mere 11 bytes can trigger the allocation of up to 131 KB of memory for a message that never arrives. This tiny trigger can bring a server to its knees, freezing memory and blocking critical connections.

Analyst 207
Dimly lit server room with one server showing high memory usage.

OpenSSL Servers Vulnerable to Memory-Bloating DDoS Attacks

Beware: a simple 11-byte malicious input can cripple OpenSSL servers with a devastating DDoS attack, leaving them permanently bloated and vulnerable. This sneaky exploit, dubbed HollowByte, takes advantage of a weakness in OpenSSL's TLS handshake to drain server resources.

Analyst 207