The Agent Skills Directory

[Indirect Prompt Injection] (HIGH): The skill processes mathematical expressions provided as strings (e.g., 'cos(w*t)') for time-dependent Hamiltonians in solvers like sesolve and mesolve. If an agent extracts these strings from untrusted sources (web scraping, user chat), it creates a vector for code injection during the runtime compilation phase.
Ingestion points: String arguments for Hamiltonian definitions in references/time_evolution.md.
Boundary markers: Absent; the skill does not suggest delimiters or 'ignore' instructions for mathematical input.
Capability inventory: Runtime compilation of strings into executable code via QuTiP's internal mechanisms.
Sanitization: Absent; no mention of validating or sanitizing mathematical strings before processing.
[Remote Code Execution] (HIGH): The Result.load() function in references/time_evolution.md is likely based on Python's pickle module, which is inherently unsafe. An attacker could provide a malicious .dat file that executes arbitrary code when loaded by the agent.
[Dynamic Execution] (MEDIUM): The skill emphasizes the use of 'string-based' time dependence for performance, which involves dynamic code generation and compilation at runtime.
[External Downloads] (LOW): The skill recommends installing qutip, qutip-qip, and qutip-qtrl. While these are legitimate scientific libraries, they are not on the pre-approved 'Trusted Sources' list provided in the analysis framework.

qutip