Skip to main content
Palyra employs a comprehensive fuzzing strategy to ensure the robustness and security of its critical parsing and data-handling components. Given the system’s reliance on processing untrusted input from LLMs, external webhooks, and workspace files, the fuzzing harness is designed to identify edge cases, memory safety issues, and logic errors before they can be exploited. The fuzzing infrastructure is built using cargo-fuzz and libfuzzer-sys, targeting various subsystems including configuration parsers, protocol decoders, and redaction routines fuzz/Cargo.toml#1-11.

​
Fuzzing Architecture and Data Flow

The fuzzing harness operates by providing pseudo-random byte streams to specialized “fuzz targets.” These targets map the raw input to specific internal functions, often with size constraints to prevent resource exhaustion during campaigns.

​
System-to-Code Mapping: Fuzz Targets

Fuzz Target NameCode Entity TargetedPurpose
config_path_parserConfigPathValidates file path resolution and sandbox escape prevention in configuration.
a2ui_json_parserpalyra_a2uiTests the resilience of the Agent-to-User Interface JSON Patch parser.
webhook_payload_parserWebhookEnvelopeEnsures external webhook signatures and payloads are handled safely.
workspace_patch_parserapply_workspace_patchValidates the logic for applying file diffs within the workspace sandbox.
process_runner_input_parserparse_process_runner_tool_inputHardens the parser for tool execution payloads.
redaction_routinesredaction.rsEnsures sensitive data (keys, tokens) is correctly identified and masked.
webhook_replay_verifierWebhookReplayTests protection against replay attacks for external triggers.
Sources: fuzz/Cargo.toml#18-80

​
Fuzzing Data Flow Diagram

The following diagram illustrates how raw fuzzer input is transformed and routed to the internal logic of the palyra-common and palyra-daemon components. “Fuzzing Data Flow” Sources: fuzz/fuzz_targets/process_runner_input_parser.rs#8-13, fuzz/fuzz_targets/workspace_patch_parser.rs#21-41, crates/palyra-common/src/process_runner_input.rs#26-31, crates/palyra-common/src/redaction.rs#31-63

​
Key Fuzzing Targets

​
Workspace Patch Parser

The workspace_patch_parser targets the apply_workspace_patch function. This is critical because it handles file modifications requested by agents. The fuzzer ensures that even with malicious patch data, the parser respects WorkspacePatchLimits, such as max_patch_bytes (16 KB) and max_files_touched (32) fuzz/fuzz_targets/workspace_patch_parser.rs#10-39.

​
Process Runner Input

The process_runner_input_parser targets the tool input for palyra.process.run. It uses serde_json with deny_unknown_fields to ensure strict schema adherence crates/palyra-common/src/process_runner_input.rs#5-7. The fuzzer attempts to break the parser by providing malformed JSON or unexpected types to the ProcessRunnerToolInput struct fuzz/fuzz_targets/process_runner_input_parser.rs#8-13.

​
Redaction Routines

The redaction_routines target exercises the logic in crates/palyra-common/src/redaction.rs. This subsystem is responsible for scrubbing secrets from logs and UI outputs. The redaction engine uses SENSITIVE_KEY_MARKERS to identify potential secrets: “Redaction Logic Mapping” Sources: crates/palyra-common/src/redaction.rs#25-85

​
Running Fuzz Campaigns

Fuzzing is performed using the cargo fuzz toolchain. Campaigns should be run periodically or after significant changes to parsing logic.

​
Prerequisites

  • Nightly Rust compiler.
  • cargo-fuzz installed (cargo install cargo-fuzz).

​
Execution Commands

To run a specific target (e.g., the process runner input parser): 
cargo +nightly fuzz run process_runner_input_parser
To run the workspace patch fuzzer with a timeout: 
cargo +nightly fuzz run workspace_patch_parser -- -max_total_time=3600
Sources: fuzz/Cargo.toml#1-11, fuzz/fuzz_targets/workspace_patch_parser.rs#1-10