XNews

AI safety debate: Hugging Face hack and agent risks

AI safety debate: Hugging Face hack and agent risks

US Treasury Secretary Bessent stated the government will not grant liability waivers to AI labs, referencing the Hugging Face hack and extinction risk claims. Databricks CEO Ali Ghodsi argued AI survival risk is near zero, with real threats from cyberattacks. Andrew Ng also discounted existential risk, calling for engineering fixes.

Key insights

US Treasury rejects liability waivers for AI labs

Treasury Secretary Bessent said on CNBC that the Trump administration will not provide federal liability waivers. He criticized labs claiming 10% extinction probability while seeking liability protection, and warned AI labs must disclose risks in IPO filings. He also supported open-source AI development to avoid regulatory capture.

Databricks CEO: survival risk 'near zero', real threat is cyberattacks

Ali Ghodsi argued that AI existential risk is overblown, citing four RSI criteria that are not met. He pointed to the drastic reduction in vulnerability weaponization time (from years to hours) as the true risk, and warned that enterprise AI adoption is bottlenecked by missing organizational ontology, not model intelligence.

Andrew Ng: Hugging Face incident due to bugs, not AI magic

Andrew Ng wrote that the Hugging Face hack was enabled by faulty sandboxing and monitoring, not superintelligence. He noted that parallel agent processes are normal, and responsibility lies with builders, not tools. He opposed pausing AI, arguing it would delay fixing safety engineering problems and harm competitiveness.

Criticism of 'pacing the frontier' framing

Martin Casado argued that Anthropic's 'pace the frontier' call was a PR mistake: safety is orthogonal to speed, labs lack credibility, and it falls into a 'valley of terror' that satisfies neither side. Ali Ghodsi added that no lab will unilaterally slow down due to competition and IPO pressures, making government regulation a forced collective brake.

Real risks: AI-driven cyberattacks and security automation imperative

Ali Ghodsi highlighted that vulnerability weaponization time has dropped to hours, and human response is insufficient. He urged fully automated detection and response for SOCs. Andrew Ng acknowledged the changed cyber threat landscape but noted that attackers still face bottlenecks, and defenders ultimately have the advantage.

Enterprise AI adoption bottleneck: missing organizational context

Ali Ghodsi said current enterprise AI use is like a 'glorified Google search' due to lack of whole-enterprise context. He advocated building an ontology (knowledge graph) from meeting transcripts and content, analogous to an offline PageRank index. Databricks itself built such an ontology with millions of nodes, transforming their operations.

Source posts