CATLASS Repository Security Statement

System Security Hardening

You can enable ASLR (Level 2) during system configuration to enhance system security and ensure system randomization is activated. Refer to the following method for configuration:
```
echo 2 > /proc/sys/kernel/randomize_va_space
```

To ensure security and minimize permissions, you are not advised to use administrator accounts such as root.

You are advised to set the umask value of the running system to 0027 or higher on the host (including the host) and container, to ensure that the default highest permission of new folders is 750 and that of new files is 640.
You are advised to take security measures such as permission control on sensitive content, including personal privacy data, business assets, source files, and various files saved during operator development. For example, control the permissions of the installation directory related to the CATLASS repository and the public data files input to operators. The recommended permissions can be set with reference to [Appendix A Recommended Maximum Permission Control Value for Files (Folders) in Each Scenario](#a-Recommended Maximum Permission Control Value for Folders in Each Scenario).
During operator runtime, compiled operator files may be cached in the kernel_meta_* folder under the running directory to accelerate subsequent operator calls. You can independently control the permissions of the generated files as needed.
You must implement proper permission control during the installation and usage process, with reference to the file permission guidelines in [Appendix A Recommended Maximum Permission Control Value for Files (Folders) in Each Scenario](#a-Recommended Maximum Permission Control Value for Folders in Each Scenario).

When compiling and running template library examples, some intermediate files will be generated. It is recommended to set appropriate permissions for these intermediate files after compilation to ensure file security.

You are advised to write an operator calling script based on the operating environment resources. If the operator calling script does not match the resource status, for example, the space used for generating input data and benchmark computing results exceeds the memory capacity limit, or the data stored locally in the script exceeds the disk space, an error may occur and the process may exit unexpectedly.
If an operator exits abnormally during runtime, it will terminate the process and print error information. It is recommended to locate the specific cause of the error based on the error prompts, including enabling the operator synchronous execution function and checking CANN log files.
When operators are called using PyTorch, PyTorch is used. This may result in runtime errors due to version mismatch. For details, see PyTorch Security Statement.

CATLASS sample aims to provide a minimal implementation for quickly getting started with, developing, and debugging CATLASS operators. Its core objective is to demonstrate CATLASS's core functionality using the simplest possible code, rather than providing production-level security. Compared to mature inference frameworks, the security features (such as input validation and boundary checks) in this sample are relatively limited.
CATLASS does not recommend that you directly use the sample as the service code, and does not ensure the security of such practices. If you directly apply these sample codes to their real business scenarios and any related security issues occur, the responsibility shall be borne solely by you.

As an NPU operator template library, CATLASS's core responsibility is to provide high-performance and functionally correct operator template implementations (located in include). To pursue optimal NPU performance, the library implementation includes (but is not limited to) the following design trade-offs:
- No address validation for buffer (e.g., UB/L1/L0) accesses
- No validation of the representation range for large-to-small type conversions
- No division-by-zero validation for division operations (e.g., Scalar / or Vector Div)
- No validation of the validity of loop conditions passed from external sources
- No validation to prevent overflow of arithmetic operation results
This design is based on the principle of performance priority: NPU validation relies on scalar units, and additional security instructions will cause performance degradation, so runtime security checks are intentionally omitted.
CATLASS only guarantees the functional correctness of operator templates under valid inputs. Any security risks arising from invalid inputs (e.g., zero/overflow values as divisors, illegal memory accesses, or unmet boundary conditions) shall be borne by the caller.

Currently, the CATLASS repository code does not involve public network addresses.

Type	Maximum Linux Permission
Home directory	750 (rwxr-x---)
Program files (including scripts and libraries)	550 (r-xr-x---)
Program file directory	550 (r-xr-x---)
Configuration files	640 (rw-r-----)
Configuration file directory	750 (rwxr-x---)
Log files (recorded or archived)	440 (r--r-----)
Log files (being recorded)	640 (rw-r-----)
Log file directory	750 (rwxr-x---)
Debug files	640 (rw-r-----)
Debug file directory	750 (rwxr-x---)
Temporary file directory	750 (rwxr-x---)
Maintenance and upgrade file directory	770 (rwxrwx---)
Service data files	640 (rw-r-----)
Service data file directory	750 (rwxr-x---)
Key components, private keys, certificates, and ciphertext file directory	700 (rwx------)
Key components, private keys, certificates, and ciphertext files	600 (rw-------)
APIs and scripts for encryption and decryption	500 (r-x------)