AI Engineering Signal #55
OpenAI ships Lockdown Mode to block prompt injection attacks on sensitive data in agentic workflows
Signals
OpenAI ships Lockdown Mode to block prompt injection attacks on sensitive data in agentic workflows
any production agent handling confidential data needs this evaluated against your threat model before the next deployment cycle.
TechCrunch
ERCOT flags data center loads tripping grid stability
procurement teams need failover power contracts before Texas capacity constraints bite.
Web
Gemma 4 31B FP8 matches Claude Sonnet 4.6 in local benchmarks
update your open-weight routing assumptions; self-hosted inference now competes at Sonnet tier.
llama.cpp merges Gemma 4 MTP support
Gemma 4 multi-token prediction is now runnable locally; retest throughput numbers on existing inference rigs.
GitHub
Attack selection in agentic AI control evals materially reduces measured safety
safety benchmarks for agents are understating risk; audit your eval harness attack coverage.
ArXiv
Notion-Anthropic service disruption resolved
single-vendor AI dependencies need documented fallback paths; this outage is the template for your runbook.
TechCrunch
The Take
The infrastructure layer — power, failover, prompt security, and eval integrity — is now the binding constraint, not model capability. Teams still optimizing for benchmark scores while skipping grid risk, injection hardening, and honest attack coverage in safety evals are building on an unstable foundation.
Subscribe
Related Signals