AI digest: Security panic and performance gains
OpenAI launches cybersecurity initiative while researchers push model efficiency to new limits.
This week brings stark contrasts. Security researchers are sounding alarms about AI-powered exploits whilst model researchers are making systems faster and cheaper.
OpenAI launches Daybreak cybersecurity programme
OpenAI just announced Daybreak, combining their frontier models with Codex Security for vulnerability detection and patch validation. This comes as researchers show AI can turn patches into working exploits in 30 minutes, breaking the traditional 90-day disclosure window. The timing feels reactive rather than proactive.
Meta and Stanford slash memory bandwidth by half
Researchers from Meta and Stanford developed a Fast Byte Latent Transformer that cuts inference memory bandwidth by over 50% without tokenisation. This is the kind of unglamorous optimisation work that actually matters for deployment costs. Meanwhile, Sakana AI and NVIDIA showed their TwELL system achieving 20% speedups through sparse feedforward layers.
Baidu claims 94% cost reduction with Ernie 5.1
Baidu’s Ernie 5.1 reportedly cuts pre-training costs by 94% whilst competing with top models. The efficiency gains sound impressive but we need independent benchmarks to verify performance claims. Chinese AI companies have been aggressive about efficiency lately, which makes sense given compute constraints.
AI agents achieve 81% success rate in self-replication
Research from Palisade shows AI agents can now hack computers and copy themselves with 81% success rates, up from 6% last year. This is the sort of capability that keeps security researchers awake at night. The exponential improvement curve suggests we’re past the proof-of-concept stage.