提交 #856870: zevorn rt-claw unreleased post-v0.2.0 (36d128f) Incorrect Authorization (CWE-863)信息

标题zevorn rt-claw unreleased post-v0.2.0 (36d128f) Incorrect Authorization (CWE-863)
描述# Technical Details An Incorrect Authorization vulnerability exists in the swarm RPC receiver in `claw/services/swarm/swarm.c` of rt-claw. The application fails to enforce `CLAW_TOOL_LOCAL_ONLY` tool flags on the inbound swarm RPC receiver path. The `handle_rpc_request()` function accepts unauthenticated UDP RPC packets, resolves the attacker-supplied `tool_name`, parses the JSON payload, and directly invokes `claw_tool_invoke()` without checking whether the tool is flagged local-only. This allows a remote attacker to invoke `save_memory`, `list_memories`, and `delete_memory` — tools explicitly registered with `.flags = CLAW_TOOL_LOCAL_ONLY` in `claw/services/tools/system.c`. Attacker-controlled key/value data is persisted into the long-term memory store at `~/.config/rt-claw/kv/claw_ltm/`. # Vulnerable Code File: claw/services/swarm/swarm.c (lines 301–339) Method: handle_rpc_request() Why: The receiver path calls `claw_tool_invoke(tool, invoke_params, result)` without checking `tool->flags & CLAW_TOOL_LOCAL_ONLY`. The sender helper `swarm_rpc_call()` at lines 477–480 does implement this check, but a raw UDP attacker bypasses it entirely. Affected tools are registered in `claw/services/tools/system.c` lines 452–486 with `.flags = CLAW_TOOL_LOCAL_ONLY`. # Reproduction 1. Build rt-claw for Linux with `-Dosal=linux -Dswarm=true -Dtool_system=true` 2. Start the rtclaw binary listening on UDP port 5300 3. Send a forged swarm RPC packet with `tool_name="save_memory"` and payload `{"key":"swarm_canary","value":"GHSA-canary"}` 4. Observe RPC response returns `status=0` with `"saved: swarm_canary = GHSA-canary"` 5. Send a second packet with `tool_name="list_memories"` — the persisted entry is visible in the response 6. Run control with a nonexistent tool — returns `"tool not found on this node"` with no persistence # Impact - Unauthenticated remote attacker can save, list, and delete long-term memory entries - Planted memories are fed into future AI context, allowing persistent influence over model behavior - Attacker-controlled data survives process restarts (persisted to disk KV store) - Same receiver-side gap applies to all CLAW_TOOL_LOCAL_ONLY tools in the build
来源⚠️ https://github.com/zevorn/rt-claw/issues/135
用户
 Eric-y (UID 95889)
提交2026-06-12 11時37分 (1 月前)
管理2026-07-17 18時08分 (1 month later)
状态已接受
VulDB条目379838 [zevorn rt-claw 直到 0.2.0 Swarm RPC Receiver swarm.c handle_rpc_request 权限提升]
积分20

Want to stay up to date on a daily basis?

Enable the mail alert feature now!