Fix detector bypass regressions in CI/CD and markdown image exfiltration rules

[cdayAI]

Motivation

• A recent change narrowed data-exfiltration parameter detection and added a broad negation lookahead that allowed crafted CI/CD @agent prompts and markdown-image URLs to bypass detection.

Description

• Restored markdown-image exfiltration coverage to include key= in both legacy and v14 data-exfiltration regexes in python-sdk/agent_shield/detector.py so payloads like ?key=SECRET... are detected.
• Adjusted the CI/CD @agent exfiltration regex in python-sdk/agent_shield/detector.py so an early benign-looking negation no longer suppresses a later explicit exfiltration instruction while preserving the narrow exception for the common benign phrase leak any sensitive data.
• Added two regression tests to python-sdk/tests/test_detector.py that assert detection for the negation-prefix CI/CD bypass and a ?key= markdown image exfil payload.

Testing

• Ran the v14 category tests with python -m pytest tests/test_detector.py -q in python-sdk and all tests passed (53 passed).
• Existing detector unit tests that exercise related categories were executed as part of the same test run and succeeded.

---

Codex Task