The Aikido Machine: on-prem AI pentesting that never leaves your network

Deel's whole job is making it possible to hire flexibly: whether the people are remote, contractors, or spread across entities and countries you don't personally manage. That flexibility is exactly what makes security hard to keep consistent. Access often outlives the project it was granted for. A package, browser extension or AI skill gets installed on a whim. A new dependency reaches a repo before anyone's reviewed it.

Altar is Aikido's open-weight security model, derived from GLM-5.3 via quantization and expert pruning (1.51 TB to 328 GB), running on-prem in Aikido Machine, including air-gapped networks.

Aikido recaps Jason Haddix's Secure Disclosure episode: he predicts 90% of pentests will be done by AI, argues defenders must fight AI with AI, and says AI solves pentesting's coverage problem.

YesWeHack examines three key AI pentesting concerns - scope control, harm avoidance, and traceability - and shows how its Agentic Pentest harness addresses them for safe autonomous testing.

GitLab Duo Self-Hosted can route through Privatemode AI: prompts, source code and completions stay encrypted end-to-end on confidential-computing hardware—no GPUs, no third party can read the code.

Wiz Research analyzes NordStellar logs to map credentials harvested by infostealers like Lumma and RedLine, showing how stolen keys and tokens unlock AWS, Azure, GitHub, and AI platforms.