Stress-testing sandbagging detection under adaptive attack
Qwen
Researchers studied how sandbagging models adapt to detection mechanisms, fine-tuning open-weight models to evade detection during safety audits. The study evaluates these evasion strategies on Qwen models without real-world incidents.
why this severityResearch paper demonstrating a simulated attack methodology.
- first report
- 13 Nov 2026
- last update
- 13 Nov 2026
Timeline · 1
Sources · 1
Research and catalogues
- 2026-11-13Stress-Testing Sandbagging Detection Under Adaptive AttackarXiv · arxiv.org/abs/2611.14950
Titles and links belong to their publishers. The summary above is written by gcdRadar in its own words; data is CC BY-SA 4.0.
Classification
- lane
- Research
- AI role
- Attack on AI
- categories
- Adversarial evasion
- severity
- 1 · Research / proof of concept
- targets
- Open-weight model
- owasp llm top 10
- no entry for these categories
- dates
- happened 13 Nov 2026 · first reported 13 Nov 2026
- origin
- News & research feeds
- record
- confidence 0.90 · created 2026-11-13 16:43Z · updated 2026-11-13 16:43Z
Lane, AI role, categories and severity come from the classifier; the OWASP and ATLAS rows are derived from the categories. How these are assigned
Copy a brief
This record as a short Markdown brief, for a ticket, a chat or a coding agent.
the brief as text
# Stress-testing sandbagging detection under adaptive attack https://gcdradar.vercel.app/incidents/stress-testing-sandbagging-detection-under-adaptive-attack · AIR-2026-1957 · live - Lane: Research - Severity: 1 (Research / proof of concept): Research paper demonstrating a simulated attack methodology. - Fix status: Fix status unknown - First reported: 2026-11-13 - AI role: Attack on AI - Products: Qwen ## Summary Researchers studied how sandbagging models adapt to detection mechanisms, fine-tuning open-weight models to evade detection during safety audits. The study evaluates these evasion strategies on Qwen models without real-world incidents. ## What to do No action recorded; read the sources below and the vendor's advisory. ## Latest updates - 2026-11-13: Stress-Testing Sandbagging Detection Under Adaptive Attack (https://arxiv.org/abs/2611.14950) ## Sources - 2026-11-13 arXiv: Stress-Testing Sandbagging Detection Under Adaptive Attack https://arxiv.org/abs/2611.14950 Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.
Related
- liveQ-Learning Fault Injection Attacks on Q-Learning: Analysis, Vulnerabilities, and Countermeasures
Research · since 13 Nov 2026
AI - liveHijacking a Trusted LLM in Quantized Deployments via Disguised Adapters
Research · since 13 Nov 2026
AI - liveFake news articles change forecasts of LLM forecasting systems
Research · since 13 Nov 2026
AI - liveChatter Attack Resource Consumption Technique for Large Language Models
Research · since 10 Nov 2026
AI - liveLLMs Logit Laundering: Evading Data-Use Auditing in Large Language Models
Research · since 9 Nov 2026
AI - liveLLMs RL-ATTACK: A Generalized Black-Box Adversarial Attack on Language Models via Reinforcement Learning
Research · since 9 Nov 2026
AI