Skip to content

Blue patch: Add generator replay detection (Pattern 16) #37

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
lalith-AE30 wants to merge 1 commit into from
Closed

Blue patch: Add generator replay detection (Pattern 16) #37

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 25 additions & 0 deletions kernelguard.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,6 +180,9 @@
RE_CPP_MARKERS = re.compile(r'#include|__global__|__device__|__host__|extern\s+"C"|asm\s+volatile')
RE_TRIPLE_QUOTED = re.compile(r'(?:r)?(?:"""|\'\'\')(.*?)(?:"""|\'\'\')', re.DOTALL)

# Pattern 16: Generator-based replay attack
RE_GEN_REPLAY = re.compile(r'def\s+\w+\s*\([^)]*\)\s*:[^}]*yield[^}]*next\s*\([^)]*\)\s*[^=]*=\s*\w+\.send')


# ---------------------------------------------------------------------------
# Utility: strip C++/CUDA inline source from Python code
Expand Down Expand Up @@ -1397,6 +1400,26 @@ def detect_token_paste_cuda_api(code: str | SubmissionFacts) -> list[dict]:
}]


def detect_generator_replay(code: str | SubmissionFacts) -> list[dict]:
"""Pattern 16: Generator-based replay attack.

Detects generators whose .send() method is exposed as custom_kernel.
The attack: create generator, prime with next(), assign g.send to custom_kernel.
"""
facts = ensure_submission_facts(code)
raw = facts.raw_code
matches = []

if RE_GEN_REPLAY.search(raw):
matches.append({
"pattern": "GENERATOR_REPLAY",
"severity": "high",
"evidence": "Generator-based replay attack: generator.send() exposed as callable",
})

return matches


# Sequence-aware batch graph exploit: learns repeated call sequences
# and builds a batch CUDA graph that launches all calls at once.
RE_CALL_SEQUENCE = re.compile(r'\b_?call_sequence\b|\b_?known_seq\b|\bbatch_graph\b|\bbatch_handle\b')
Expand Down Expand Up @@ -3773,6 +3796,7 @@ def support_only_patterns(matched_patterns: list[dict]) -> bool:
detect_unsync_multistream,
detect_cuda_event_disable_timing,
detect_token_paste_cuda_api,
detect_generator_replay,
detect_sequence_batch_graph,
detect_runtime_package_install,
# AST-based detectors (Layer 2)
Expand Down Expand Up @@ -3811,6 +3835,7 @@ def support_only_patterns(matched_patterns: list[dict]) -> bool:
("unsync_multistream", detect_unsync_multistream),
("cuda_event_disable_timing", detect_cuda_event_disable_timing),
("token_paste_cuda_api", detect_token_paste_cuda_api),
("generator_replay", detect_generator_replay),
("sequence_batch_graph", detect_sequence_batch_graph),
("runtime_package_install", detect_runtime_package_install),
("trusted_module_import", detect_trusted_module_import),
Expand Down