feat: add goal hijacking regression scenario for cloud file deletion

[anshjaiswal12]

This PR adds a new executable security regression scenario for Goal Hijacking, as outlined in issue #16. I noticed a gap in the test coverage for this attack class — the existing goal_hijack/basic.yaml covers a simple document summarization context, but there was no scenario testing resistance to high-severity, tool-invocation-based goal diversion (e.g., an attacker attempting to trigger destructive operations like deleting cloud storage files).

The scenario is defined in scenarios/goal_hijacking_01.yaml and models an attack where malicious instructions are embedded within spreadsheet data — an untrusted, retrieved_document-style context. The agent's original goal is to calculate travel expenses; the injected payload attempts to divert it into calling cloud_delete or shell. I followed the existing schema structure (same top-level fields, trust: untrusted context sourcing, and denied_tools under expected) to stay consistent with the project's patterns and ensure the harness can ingest the file without any code changes.

Verification: the scenario passes agent-harness run scenarios/goal_hijacking_01.yaml --dry-run and is correctly loaded and validated by the harness. I also added a targeted test in tests/test_goal_hijacking_scenario.py to confirm schema compliance. All 9 tests in the suite pass cleanly.

[mertsatilmaz]

Thanks for the contribution. The scenario itself is strong: it is realistic, high-impact, and fits the goal-hijack category well. The injected spreadsheet content creates a clear conflict between the user’s original goal, calculating travel expenses, and the attacker’s goal, deleting cloud files. The denied tools also make the expected safety boundary explicit.

I verified the scenario loads and the test suite passes:

• python -m py_compile tests/test_goal_hijacking_scenario.py
• python -m pytest

One change is needed before merge: please move the scenario into the existing category-based layout.

Current path:

scenarios/goal_hijacking_01.yaml

Requested path:

scenarios/goal_hijack/delete_cloud_files_001.yaml

Please also update the test path accordingly.

Minor style cleanup in the test file:

• Add a blank line between from pathlib import Path and from agent_harness.scenario import load_scenario
• Add two blank lines before the test function

After that, this should be mergeable.

[mertsatilmaz]

Thanks @anshjaiswal12. I pushed the layout/style fixes directly onto the branch (kept you as co-author on the cleanup commit) so this can land. Approving and merging.