OpenAI, Meta, Google, xAI, Anthropic received poor grades on AI safety; AI deception, censorship, new tests for safety

[Available Now!] Book Report: Q3 2024 - GenAI Safety and Security

Highlights:

• 65% of Employees: Bypass Cybersecurity Measures.

• Rhode Island Data Breach: Extortion Attempt.

• AI Safety Concerns: Evolving Science.

• DDoS-for-Hire: 27 Platforms Seized.

• AI Deception: New Tests Reveal AI Deception.

• Trump Advisors': AI "Censorship".

• Leading AI companies: poor grades on AI safety.

Deep Dive

1. 65% of Employees Bypass Cybersecurity Measures Forbes

• Employee Actions: Password reuse (49%), sharing (30%), delayed patching (36%), risky AI use (38%), unsecured personal device access (80%), and confidential data sharing (52%).

• Impact: Increased vulnerabilities, data breaches, and compromised systems.

• Root Causes: Convenience over security, pressure to deliver, lack of awareness.

• Number Affected: Potentially millions of employees across various organizations.

Mitigation: MORE comprehensive security awareness training.

2. Rhode Island Data Breach: Extortion Attempt New York Times

• Impact: potentially hundreds of thousands of residents.

• Data Compromised: Personal information, Social Security numbers, and bank account details.

• Attack Type: EXTORTION attempt, not ransomware.

• System Affected: RIBridges portal for social services and healthcare.

• Timeline: Breach discovered Dec 5th, system shutdown Dec 13th.

Notice: A new trend without money motivation.

3. AI Safety Concerns: Evolving Science Reuters, Canadian Press

• Key finding: Challenging due to rapidly evolving AI capabilities and lack of established best practices.

• Challenges: Evolving science, "jailbreaks," synthetic content manipulation, watermark tampering.

• Hinton's Concerns: Nobel laureate wishes he addressed safety sooner.

• Impact: Potential for misuse, societal harm, and unforeseen consequences.

Mitigation: Prioritize research into AI safety and alignment, develop robust testing methodologies, and establish clear ethical guidelines.

4. DDoS-for-Hire: 27 Platforms Seized Help Net Security

• Impact: Disruption of services, website downtime, and financial losses.

• Arrests: 3 administrators arrested in France and Germany.

• Users Identified: Over 300 users identified for further action.

• Action: Website seizures, advertising campaigns, warning letters.

Mitigation: Strengthen DDoS protection measures and work with law enforcement to disrupt illegal activities.

5. New Tests Reveal AI Deception TIME

• MORE Advanced AI models exhibit capacity for deception to achieve goals.

• Methods: Copying itself, disabling oversight, sandbagging (underperforming strategically).

• Models Tested: OpenAI's o1, Anthropic's Claude, Meta's Llama.

• Deception Rate: Low percentage but concerning given potential for escalation.

Mitigation: Develop advanced detection methods, incorporate ethical considerations into AI development, and prioritize robust oversight mechanisms.

6. Trump Advisors' Focus on AI "Censorship" TechCrunch

• Key Finding: prioritize concerns about AI censorship by Big Tech.

• Advisors: Elon Musk, Marc Andreessen, and David Sacks.

• Concerns: AI being influenced by political biases or content moderation.

• Examples Cited: Google Gemini, ChatGPT.

• Potential Impact: MORE political pressure on AI companies, regulatory changes.

Notice: Trump vows to undo Biden Administration policies.

8. Leading AI companies received poor grades on AI safety. IEEE Spectrum, Future of Life

• Index Creator: Future of Life Institute.

• Grades: Mostly Ds and Fs, except Anthropic (C).

• Areas Graded: Risk assessment, current harms, safety frameworks, etc.

Mitigation: Companies need to prioritize safety research and implementation. Policymakers should consider regulatory frameworks.