Security & safeguards

How the skills are hardened. (For reporting a vulnerability, see the repository's SECURITY.md disclosure policy.)

All skills are hardened against the OWASP Top 10 for LLM Applications, with dedicated validation and edge-case tests. Key safeguards:

• Tool access is explicit — skills that only inspect or write files avoid Bash; script-bearing skills declare Bash and keep executable behavior inside reviewed Python CLIs.
• Input validation at every boundary — numeric parameters are bounds-checked and validated as finite; string inputs (parameter/field/term names) are validated against regex allowlists.
• Safe file loading — JSON/CSV/NPY loaders enforce file-size limits (100–500 MB) and structure validation (dict root required); np.load() uses allow_pickle=False.
• No eval()/exec() — region-condition parsing uses strict regex matching, never dynamic code execution.
• Prompt-injection resistance — string values extracted from external files are truncated and stripped of control characters before surfacing to the agent; log-derived phase names are sanitized.
• Command-construction safety — shlex.quote() escapes paths interpolated into shell commands; command templates are validated against a shell-operator denylist.
• ReDoS prevention — user-supplied regex patterns are length-capped and checked for catastrophic backtracking.

Each skill documents its specific safeguards in a ## Security section within its SKILL.md, with standardized subsections for Input Validation, File Access, Tool Restrictions, and Safety Measures — enforced by mss validate (see PROTOCOL.md).

Security risk tiers

Every skill is classified by its tool-access surface:

Tier	Criteria	Skills
HIGH	Has Bash (can execute scripts)	16 — numerical-stability, time-stepping, convergence-study, differentiation-schemes, nonlinear-solvers, ontology-explorer, ontology-validator, simulation-validator, slurm-job-script-generator, benchmark-and-mms-planner, workflow-engine-mapper, fair-simulation-packager, md-analysis-planner, hpc-runtime-doctor, simulation-failure-triage, skill-evaluator
MEDIUM	Has Write but no Bash	7 — linear-solvers, mesh-generation, numerical-integration, parameter-optimization, performance-profiling, post-processing, simulation-orchestrator
LOW	Read/Grep/Glob only	1 — ontology-mapper

Per-skill tiers are also in the machine-readable skills_index.json.