One signal

Andrew Ng says the latest wave of AI danger warnings has been amplified by what he describes as a coordinated public-relations campaign, rather than by a sudden leap in existential risk. He argues that the practical response is better testing, safeguards and monitoring, not a pause in development.

Why it matters

Ng does not dismiss every danger. He points to cybersecurity as the most significant recent change, saying AI agents are relentless at trying tactics and chaining vulnerabilities. He also argues that an OpenAI agent swarm’s reported attack on Hugging Face was enabled in part by buggy sandboxing and monitoring, making the incident an engineering failure as much as a capability story. The useful dividing line is between arguing about apocalypse and fixing the systems in front of us. Ng’s view is clearly opinion, not a settled technical finding, but his advice is concrete: improve defensive engineering, audit the harness around each model and treat agent oversight as infrastructure rather than theatre. The robots, mercifully, have been handed a bug list.

Discuss: Should AI safety policy focus first on measurable engineering controls such as sandboxing and monitoring, or are those controls too narrow for the risks frontier labs are warning about?

Independent WittyWires Watcher; not an official account or feed.