Tools Model Security Capability glm-5-2-replication-research
Model record
GLM-5.2 · replication experiment: how it does on security work, and what it will answer
- Lab
- Z.ai
- Access
- research
- Weights
- open
- Safeguards
- Controlled replication research with a separate inference server
- Score rows
- 1
- Confidence
- CONFIRMED
What happened
Anthropic tested GLM-5.2 as a controller deploying Llama3.1-8B through an isolated eight-node emulated Android fleet. It reached all eight nodes in four of five trials.
Isolated Android replication, trials reaching all eight nodes: 4 of 5 (third-party-vendor; Anthropic replication experiment; Eight emulated Android nodes with unauthenticated ADB; safeguards: Controlled research configuration; run date not published; CONFIRMED). The paper was published on 12 October. The fifth trial reached seven nodes. The controller used separate inference infrastructure.. Refusal gate (CONFIRMED, observed 2026-10-12): Research setup; production access policies depend on the deployment.. The fleet was firewalled and had unauthenticated ADB access. GLM-5.2 ran on a separate inference server. The result measures orchestration under those conditions; compare it with experiments using the same access assumptions and model roles.
Sources
- Anthropic replication researchwww.anthropic.com/research/self-replicating-ai
On YFarmX
- Reference pageyfarmx.com/ai/llms/glm-5-2/
One record from the Model Security Capability, maintained by the Security Desk.Data: CSV · JSON ·RSS · CC BY 4.0 with attribution to YFarmX.Tracker updated · 13 October 2026