AI Engineering Signal #25
Mistral Medium 3.5 128B drops as open-weight
Signals
Mistral Medium 3.5 128B drops as open-weight
a 128B model positioned explicitly around reliability over raw benchmark chasing, now on Hugging Face for self-hosting.
Web
LLM safety broken via incremental token completion
one-word-at-a-time decomposition bypasses alignment guardrails; patch your input pipelines.
ArXiv
Alignment whack-a-mole: finetuning restores copyrighted recall
safety training suppresses but doesn't erase memorized content; fine-tune at your own legal risk.
GitHub
Ramp's Sheets AI exfiltrates financials
prompt injection via spreadsheet data leaks sensitive numbers to attacker-controlled endpoints; real production incident, not a lab demo.
Web
Physicists establish universal speed limit on quantum information scrambling
bounds how fast entanglement can spread, with direct implications for quantum error correction timelines.
Web
Figure AI hits 1 robot per hour production rate
24x scale increase signals humanoid manufacturing is crossing from prototype to supply chain reality.
Web
Nature Medicine demands clinical evidence for medical AI
calls out the gap between benchmark performance and demonstrated patient outcomes; regulatory pressure incoming.
Web
The Take
Two attack vectors confirmed in production this week — incremental jailbreaks and spreadsheet prompt injection — while alignment research keeps showing safety training is a surface coat, not a substrate. The reliability problem isn't theoretical anymore; it's in your users' spreadsheets right now.
Subscribe
Related Signals