GLM-5.2 · workflow
A real, successfully run “multi-model division-of-labor audit workflow”: GLM-5.2 handled code auditing while GPT-5.5 handled context collection, scanning firmware source code and finding high- and critical-severity vulnerabilities at a cost of about ¥1,000. The author later reproduced an almost identical vulnerability set with dsh minimal mode + DeepSeek-V4-Pro-0813 (only two were missing), bringing the cost down to ¥4. This provides a reusable reference for model selection and cost planning for “LLM-assisted source code auditing.”.
A real, successfully run “multi-model division-of-labor audit workflow”: GLM-5.2 handled code auditing while GPT-5.5 handled context collection, scanning firmware source code and finding high- and critical-severity vulnerabilities at a cost of about ¥1,000. The author later reproduced an almost identical vulnerability set with dsh minimal mode + DeepSeek-V4-Pro-0813 (only two were missing), bringing the cost down to ¥4. This provides a reusable reference for model selection and cost planning for “LLM-assisted source code auditing.”
Suitable tasks: firmware/source-code vulnerability audits (finding high- and critical-severity vulnerabilities); multi-model Agent workflows that divide “code auditing + context collection”; choosing a cost-effective model combination for security-audit tasks.
Unsuitable tasks: routine iteration where cost is extremely sensitive (the same task has a much cheaper alternative: DeepSeek-V4-Pro-0813 + dsh minimal mode costs about 1/250 as much); scenarios requiring complete, reproducible prompts/scripts (the author did not disclose workflow details or prompts).
Applicable model versions: GLM-5.2 (code-audit role); GPT-5.5 (context-collection role); comparison model DeepSeek-V4-Pro-0813 (dsh minimal mode).
Applicable clients, Agents, or APIs: any Agent framework/client that supports multi-model orchestration (the author did not identify the specific implementation, which is presumed to be a self-built workflow); dsh minimal mode (comparison group).
Recommended reasoning tier and parameters: Not disclosed; for security-audit tasks, the official coding guidance recommends using GLM-5.2's high or highest thinking tier (see documents 01 and 02 in the prompts directory).
Task: Firmware source-code vulnerability audit (target: high / critical-severity vulnerabilities)
Workflow (the author's version from two months earlier):
1. Context collection (GPT-5.5): extracts, organizes, and feeds the audit model the context related to the target from the firmware source code
2. Code audit (GLM-5.2): performs the vulnerability audit based on the collected context
3. Output: a set of high- and critical-severity vulnerabilities; total cost of about ¥1,000
Comparison group (the author's retest on 8/16):
1. dsh minimal mode + deepseek-v4-pro-0813
2. Output: an almost completely identical vulnerability set (only 2 missing); cost ¥4(Note: The author did not disclose the specific prompts, audit scripts, or vulnerability details. The table above is a structured restatement of the post's information, intended to explain the division of labor and costs.)
Conclusion: GLM-5.2 delivers real output in the “code audit” role (finding high- and critical-severity vulnerabilities), but its results can be reproduced in a similar setting by the much cheaper DeepSeek-V4-Pro-0813 (with only 2 vulnerabilities missing). Therefore, GLM-5.2 is neither the only nor the cheapest choice for audit tasks.
Environment: The author's self-built workflow; environment details were not disclosed (context-window usage, API channel, or audit scope); costs are reported by the author (about ¥1,000 vs. ¥4).
Limitations: A single case involving a single target (some firmware source code); the severity of the “2 missing” vulnerabilities was not specified; this was not a controlled comparison; the specific prompts cannot be reused (the author did not make them public).
Supporting evidence: GLM-5.2's capabilities and risks on security/audit tasks are also covered in documents 02 (NIST CAISI: safeguards permit assistance with agentic vulnerability exploitation) and 09 (Hugging Face forensic use) in this directory.
“I put together a huge workflow using GLM-5.2 (code audit) + GPT-5.5 (Context collection), scanned some firmware source… This is a necessary excerpt; read the original source for full context.
“I tried dsh minimal mode + deepseek-v4-pro-0813 today and found an almost identical set of vulnerabilities (only 2 miss… This is a necessary excerpt; read the original source for full context.
X.com (Twitter), @CossackWang (C0ss4ck, a security-focused developer) · Source date: 2026-08-16 · Edited: 2026-09-20
Read the original sourceGLM-5.2
Run this guide in the environment listed above. Downloading does not transfer the template or establish model availability for your account.