Shadow Vulnerabilities in AI/ML Data Stacks - What You Don’t Know...
summary
Open-source AI stacks harbor RCE via SnakeYAML and default misconfigurations that bind management APIs to 0.0.0.0, enabling model theft, poisoning, and remote command execution in default deployments. Researchers find thousands of exposed TorchServe and Ray instances and a chain of five CVEs around Ray that attackers exploit to run wget|bash, steal DB and API tokens, and deploy crypto miners. TF Keras H5 checkpoints allow a marshal-based code execution downgrade bypass. A working defense uses eBPF runtime sandboxing to baseline library behavior and detect deviations.
AI-written summary · about
chapters (15)
- 0:00Shadow vulnerabilities in AI
- 1:25What are shadow vulnerabilities
- 4:37Silent misconfigurations in libraries
- 6:24Why AI systems are targets
- 8:33Attacks via PyTorch TorchServe
- 13:53SnakeYAML RCE in TorchServe
- 15:22Impact of TorchServe flaws
- 16:26TensorFlow Lambda layer exploits
- 23:19TensorFlow legacy format bypass
- 24:00ShadowRay AI infrastructure attack
- 27:57Ray default deployment risks
- 30:26Ginger template injection
- 31:49Protecting AI systems
- 33:55Summary and community efforts
- 34:58Q&A on detection strategies
related talks
33mShellTorch the Next Evolution in *4Shell Executions
30mThe Simple, Yet Lethal, Anatomy of an AI Attack
56mHidden Risks in Open-Source Code and AI Models
45mThe Dark Side of AI: The Hidden Supply Chain Risks in Open...
27mSecurity panel: Can your platform survive AI?
36mThe Dark Side of AI: The Hidden Risks in Open-Source AI Models
28mFlawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...
50mMalice In Chains: Supply Chain Attacks Using Machine Learning Models