Eresus research, advisory, and security news
We collect writing, advisories, and current-event analysis around AI security, the MCP ecosystem, application security, and real attack chaining here.
Latest Posts
ONNX Unsafe Attribute Parsing (Arbitrary Code Execution / DoS)
A severe vulnerability detecting weaponized ONNX models exploiting legacy internal processing loops—manipulating attributes and external data extraction...
LlamaFile Binary Shell Overloading Threat
Identifies highly specific payload injections targeting LlamaFile's native Cosmopolitan Libc architecture—hijacking the dual-executable parsing...
LiteRT Delegation Core Execution Exploits
Represents critical systemic integration failures explicitly targeting localized Hardware Delegation Frameworks (Core ML / NNAPI / Hexagon), exploiting...
LiteRT Out-of-bounds Read/Write (Heap Corruption)
A severe vulnerability exploiting malformed shape descriptors inside customized operators executing natively on edge GPUs—triggering out-of-bounds...
PAIT-ONNX-200: Embedded Malicious Binary Artifacts and Polyglots in ONNX Models
Analysis of supply chain trojans within ONNX model files, exploring how adversaries embed hidden executables and polyglot payloads inside neural network protobuf structures.
PAIT-KERAS-301: Deferred Code Execution via Malicious Custom Layers in Keras Models
Analysis of runtime exploitation in Keras and HDF5 models, detailing how adversaries conceal unauthorized Python execution inside custom Layer definitions that trigger only during inference.
Keras Model Lambda Layer Suspicious Operator Detected at Model Load Time
High-severity deserialization threat identifying suspicious or highly uncharacteristic operational routines within a Keras Lambda layer payload upon...
Keras Extracted Object Hijacking & DoS Corruptions
Advanced exploitation where seemingly secure Keras architectures (.h5 or .keras) manipulate intrinsic layer attributes or geometry to crash execution...
Keras HDF5 Lambda Layer Arbitrary Code Execution
Identifies insecure legacy Keras objects (`.h5` or `.hdf5`) attempting to deserialize and execute malicious Python scripts hidden inside Lambda layer...