EresusSecurity
Rule Reference

Rule Reference

Sentinel detection rules organize model-file, prompt, agent, container, archive, secret, and supply-chain risks into searchable categories.

Categories

  • PICKLEDetects unsafe Python pickle streams, joblib artifacts, and embedded pickle payloads in ML model files.
  • TORCHDetects unsafe PyTorch checkpoint loading, TorchScript custom operators, and dynamic hub imports.
  • ONNXDetects risky ONNX graph metadata, external tensor paths, and custom operators.
  • SafetensorsDetects malformed safetensors headers, suspicious tensor metadata, and missing integrity controls.
  • GGUFDetects malformed GGUF metadata, risky chat templates, and untrusted quantized model bundles.
  • NetworkDetects model, agent, or evaluation configurations that can reach sensitive network locations.
  • SymbolsDetects dangerous exported symbols, native library load hooks, and unresolved entry points in model runtimes.
  • JITDetects unsafe just-in-time compilation, dynamic evaluation, and cache poisoning patterns in AI workloads.
  • CVEDetects known vulnerable packages, runtimes, model-serving components, and scanner advisory matches.
  • OCIDetects risky container image configuration, secrets in layers, and unpinned base images.
  • 7ZDetects traversal, decompression bombs, and unsafe links in 7z archives used for model bundles.
  • ManifestDetects unsafe model, plugin, MCP, and agent manifests with risky permissions or missing integrity fields.
  • Jinja2Detects server-side template injection and unsafe template constructs in prompt, chat, and report templates.
  • CompressedDetects unsafe compressed archives across ZIP, tar, gzip, xz, and nested model bundles.
  • RARDetects traversal, decompression bombs, and opaque encrypted payloads in RAR archives.
  • SecretsDetects credentials, private keys, tokens, and sensitive data embedded in model artifacts and adjacent files.
  • Size limitsReports files, archives, or tensor metadata that exceed safe scanning and runtime thresholds.

Which rule family should you start with?

If you are starting AI security from scratch, keep the order simple: first Secrets and Pickle/Torch findings with direct production risk, then Manifest/Network boundaries for agents and MCP, and finally Size limits and metadata hygiene.

RECOMMENDED START
  • If you download models or HuggingFace repos: PICKLE, TORCH, SAFETENSORS, GGUF.
  • If you use LLM agents or MCP: MANIFEST, Network, Jinja2, Secrets.
  • If you build CI/CD gates: CVE, OCI, Compressed, Size limits.

Rule index

Rule IDCategorySeverityTitleCWEFix Hint
PICKLE-EXECPICKLECRITICALDangerous Pickle ExecutionCWE-502Do not load untrusted pickle files. Convert the artifact to a non-executable format.
PICKLE-GLOBAL-IMPORTPICKLEHIGHUnexpected Global ImportCWE-502CWE-829Restrict allowed globals and require signed model artifacts.
PICKLE-STRUCTPICKLEHIGHPickle Opcode Structural TamperingCWE-915Reject structurally abnormal pickle artifacts during intake.
TORCH-LOAD-UNTRUSTEDTORCHCRITICALUntrusted torch.load DeserializationCWE-502Load tensors, not executable Python objects.
TORCH-SCRIPT-CUSTOM-OPTORCHHIGHTorchScript Custom OperatorCWE-94CWE-829Review and sign custom operators before promotion.
TORCH-HUB-DYNAMIC-IMPORTTORCHMEDIUMDynamic Torch Hub ImportCWE-829CWE-494Replace remote hub execution with pinned, vendored model code.
ONNX-EXTERNAL-DATA-PATHONNXHIGHExternal Data Path EscapeCWE-22Reject external data paths outside the model bundle.
ONNX-CUSTOM-OPONNXHIGHONNX Custom OperatorCWE-94CWE-829Approve custom operators explicitly or remove them from the graph.
ONNX-METADATA-TAMPERONNXMEDIUMSuspicious ONNX MetadataCWE-345Add provenance metadata and sign the model release.
SAFETENSORS-HEADER-OVERFLOWSAFETENSORSHIGHOversized Safetensors HeaderCWE-400CWE-770Keep safetensors headers small and predictable.
SAFETENSORS-DTYPE-SHAPE-MISMATCHSAFETENSORSMEDIUMTensor Shape and Byte Range MismatchCWE-20Validate tensor byte ranges before loading.
SAFETENSORS-MISSING-HASHSAFETENSORSLOWMissing Safetensors Integrity HashCWE-345Hash and sign safetensors artifacts.
GGUF-METADATA-OVERFLOWGGUFHIGHGGUF Metadata Overflow PatternCWE-190CWE-787Do not load malformed GGUF metadata.
GGUF-JINJA-TEMPLATEGGUFHIGHUnsafe GGUF Chat TemplateCWE-94Review tokenizer templates before use.
GGUF-UNTRUSTED-QUANTGGUFMEDIUMUntrusted Quantized ModelCWE-345Treat quantized models as release artifacts with provenance.
NET-METADATA-ENDPOINTNETHIGHCloud Metadata Endpoint AccessCWE-918Deny metadata service access from model and agent runtimes.
NET-PRIVATE-RANGE-EGRESSNETMEDIUMPrivate Network EgressCWE-200Constrain agent network access to approved hosts.
NET-UNPINNED-MODEL-URLNETMEDIUMUnpinned Remote Model URLCWE-494Download only hash-verified model artifacts.
SYM-DANGEROUS-EXPORTSYMHIGHDangerous Native ExportCWE-749Block native plugins with dangerous exports unless explicitly approved.
SYM-NATIVE-LOAD-HOOKSYMHIGHUnsafe Native Load HookCWE-829Pin native loader paths to signed artifacts.
SYM-UNRESOLVED-ENTRYPOINTSYMLOWUnresolved Native EntrypointCWE-477Keep native runtime ABI and extension versions aligned.
JIT-EVAL-COMPILEJITHIGHDynamic eval or compileCWE-95Do not compile user-controlled strings.
JIT-DYNAMIC-KERNELJITHIGHDynamic Kernel GenerationCWE-94Separate kernel code from runtime data.
JIT-CACHE-POISONINGJITMEDIUMWritable JIT CacheCWE-345Protect compiled caches from untrusted writes.
CVE-KNOWN-EXPLOITEDCVECRITICALKnown Exploited VulnerabilityCWE-1104Upgrade to a fixed version before release.
CVE-VULNERABLE-RUNTIMECVEHIGHVulnerable AI Runtime ComponentCWE-1104Patch model-serving and parser dependencies.
CVE-STALE-ADVISORY-DATACVEINFOStale Advisory DataCWE-1104Keep vulnerability intelligence current.
OCI-ROOT-USEROCIMEDIUMContainer Runs as RootCWE-250Set a non-root USER in the image.
OCI-SECRET-IN-LAYEROCICRITICALSecret in Container LayerCWE-798Never bake credentials into images.
OCI-UNPINNED-BASEOCIMEDIUMUnpinned Base ImageCWE-494Use immutable OCI digests for base images.
7Z-PATH-TRAVERSAL7ZHIGH7z Path TraversalCWE-22Block archive entries that resolve outside the destination root.
7Z-BOMB7ZHIGH7z Decompression BombCWE-409Cap archive extraction size and nesting depth.
7Z-SYMLINK-ESCAPE7ZMEDIUM7z Symlink EscapeCWE-59Treat archive links as untrusted paths.
MANIFEST-REMOTE-CODEMANIFESTHIGHManifest Loads Remote CodeCWE-829Do not execute mutable remote manifest code.
MANIFEST-OVERBROAD-TOOLSMANIFESTHIGHOverbroad Agent Tool PermissionsCWE-266Narrow agent tool permissions before deployment.
MANIFEST-MISSING-INTEGRITYMANIFESTMEDIUMMissing Manifest IntegrityCWE-353Record and verify artifact integrity in manifests.
JINJA2-SSTI-001JINJA2CRITICALServer-Side Template InjectionCWE-94Treat prompt templates as code and review them before load.
JINJA2-UNSANDBOXED-TEMPLATEJINJA2HIGHUnsandboxed Jinja2 RenderingCWE-94Sandbox template rendering and avoid dynamic template strings.
JINJA2-SECRET-EXPOSUREJINJA2HIGHTemplate Secret ExposureCWE-200Keep credentials out of render context.
COMPRESSED-ZIP-SLIPCOMPRESSEDHIGHCompressed Archive Path TraversalCWE-22Normalize archive paths before extraction.
COMPRESSED-DECOMPRESSION-BOMBCOMPRESSEDHIGHDecompression BombCWE-409Limit archive expansion and file counts.
COMPRESSED-NESTED-DEPTHCOMPRESSEDMEDIUMExcessive Nested Archive DepthCWE-400Avoid deeply nested model bundles.
RAR-PATH-TRAVERSALRARHIGHRAR Path TraversalCWE-22Block RAR entries outside the target directory.
RAR-BOMBRARHIGHRAR Decompression BombCWE-409Constrain RAR extraction size.
RAR-ENCRYPTED-PAYLOADRARMEDIUMEncrypted RAR PayloadCWE-693Do not promote opaque encrypted archives.
MODEL-SECRET-API-KEYMODEL-SECRETCRITICALAPI Key in Model BundleCWE-798Rotate exposed API keys immediately.
MODEL-SECRET-PRIVATE-KEYMODEL-SECRETCRITICALPrivate Key in ArtifactCWE-522Never publish private keys in model repositories.
MODEL-SECRET-PIIMODEL-SECRETHIGHSensitive Personal Data in Model ContextCWE-200Keep real user data out of model release bundles.
SCAN-SIZE-MAX-FILESCAN-SIZEINFOFile Exceeds Scan Size PolicyCWE-400Tune max file size intentionally.
SCAN-SIZE-ARCHIVE-RATIOSCAN-SIZEMEDIUMArchive Expansion Ratio Exceeds PolicyCWE-409Limit decompression ratio.
SCAN-SIZE-TENSOR-DIMENSIONSCAN-SIZEMEDIUMExcessive Tensor DimensionCWE-770Reject tensor metadata that exceeds runtime budgets.

Eresus support

Turn the finding into an action your team can actually close.

If you need exploit evidence, prioritization, remediation direction, and retesting for Sentinel rule set and release gate design, Eresus can help scope the work with your team.

Start Security Test