Multi-Layered AI Safety Threats and Response Strategies

AI 안전 | Thu Jul 09 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 5 sources

AI safety issues emerged across all fronts, from government partnership principles to prompt injection botnet attacks, academic cheating, geopolitical framing, and dual-use knowledge control.

Analysis

[OpenAI] published National Security Principles and expanded government partnerships ^[1]

• expanded the Daybreak cyber defense program
• signed Trusted Access for Cyber partnerships with Australia
• Canada
• Japan
• South Korea
• France
• Germany
• and others
• provided the GPT-Rosalind biosecurity model to U.S. and allied partners
• prohibited use for large-scale domestic surveillance
• autonomous weapons command
• and high-risk automated decision-making

[HalluSquatting researchers] disclosed a large-scale botnet attack technique using 9 AI coding tools ^[2]

• Cursor
• Gemini CLI
• Windsurf
• GitHub Copilot
• Cline
• and others are vulnerable
• preemptively predicts and registers resource identifiers that LLMs hallucinate
• pull-based attack enables large-scale infection and DDoS
• indiscriminate infection through automatic reverse shell installation

[Brown University] experienced a large-scale AI cheating incident in the ECON 1170 course ^[3]

• after introducing take-home exams
• enrollment surged from 30 to 86 students
• midterm average was 96 points
• with 40 students scoring perfect marks
• scores dropped 50% after switching to in-person final exams
• a Princeton survey also showed 29.9% of students admitted to AI cheating

[Verity Harding] published an essay warning of the dangers of AI arms race framing ^[4]

• former head of global public policy at Google DeepMind
• edited the essay collection Reframing the AI Arms Race
• concerned that U.S.-China competition framing blocks the possibility of international cooperation
• criticized the Trump administration's export controls and nationalist rhetoric

[Anthropic & AE Studio] released GRAM research, an off-switch technique for dual-use knowledge ^[5]

• Gradient-Routed Auxiliary Modules approach
• adds removable category-specific modules to each Transformer layer
• updates only the relevant module when training on dual-use data
• enables implementing multiple filtered versions with a single model

Sources

• [1] Our approach to government and national security partnerships - OpenAI Blog
• [2] Hackers can use 9 of the most popular AI tools to assemble massive botnets - Ars Technica AI
• [3] Suspecting AI cheating, Ivy League prof ordered in-person final; scores fell 50% - Hacker News
• [4] This Former DeepMind Exec Thinks the AI Arms Race Could End in Disaster - Wired AI
• [5] An off switch for dual use knowledge in AI models - Anthropic Research