A naukri.com initiative
Lastwatchdog
3w
112
MY TAKE: Microsoft takes ownership of AI risk — Google, Meta, Amazon, OpenAI look the other way
- At Microsoft Build, Azure CTO warned about AI vulnerabilities, including jailbreak-style attacks and structural limitations.
- Microsoft admitted Crescendo Attacks, where AI can unknowingly step over the line due to training to be helpful.
- They introduced the concept of Crescendomation, AI learning to jailbreak itself, revealing vulnerabilities.
- Microsoft acknowledged that smarter AI systems may become more vulnerable, a structural flaw in AI.
- Microsoft's Discovery platform deploys agentic AIs in R&D, addressing complex problems like designing PFAS-free cooling fluids.
- Comparison with other tech giants reveals Microsoft's transparency in addressing AI vulnerabilities, unlike Google, Meta, Amazon, and OpenAI.
- Google remains opaque about vulnerabilities in its Gemini model and focuses on performance improvements.
- Meta acknowledges safety limitations but lacks transparent articulation for hardening strategies.
- Amazon is methodical in mitigating jailbreaks but lacks public narrative, missing an opportunity to lead on trust.
- Anthropic prioritizes safety in AI models with Constitutional AI framework, emphasizing trustworthiness from the start.
Read Full Article
6 Likes
For uninterrupted reading, download the app