daily

AI Adjacent Daily Briefing – June 26, 2026

June 26, 2026

Qualcomm buys portability software, 360 makes an unverified Mythos claim, and robot sensors become prompt-injection channels.

Overview

AI competition is moving through both the software stack and the physical world. Qualcomm is paying for cross-hardware portability, China's 360 is claiming parity without a reproducible comparison, and RIPA finds prompt injection entering robots through vision, audio, and LiDAR.

1. Qualcomm agrees to buy Modular for $4 billion

Qualcomm agreed to acquire AI software startup Modular for $4 billion, Reuters reported. Modular develops software intended to make AI workloads run across different hardware, giving Qualcomm an engineering layer that could reduce dependence on ecosystems controlled by a single accelerator vendor.

The $4 billion price assigns strategic value to a software layer that can move AI workloads across accelerators. Framework compatibility, migration effort, measured end-to-end performance, and continued support for rival hardware determine whether that portability survives ownership by one chip company.

Sources: Reuters

2. China's 360 claims tools comparable with Anthropic's Mythos

Chinese technology company 360 said it developed tools capable of matching Anthropic's Mythos, Reuters reported. The statement is a company claim made amid access restrictions and intense national competition; the cited report does not provide an independent, task-by-task validation of equivalence.

The cited report supplies no independent, task-level comparison. Model version, tool access, inference budget, evaluation set, and success threshold can each change a parity result, while orchestration can make two systems appear equivalent despite different underlying model capability.

Sources: Reuters

3. RIPA carries prompt injection through robot vision, audio, and LiDAR

RIPA tests sensory prompt injection against five language models controlling a ROS 2 robot, with 100 runs per injection variant. Visual text enters through optical recognition, audio through speech transcription, and fabricated LiDAR context enters the model's system-level environment representation.

Attack success tracked model identity, not size: a 70-billion-parameter Llama and a roughly 4-billion-parameter Gemma both reached 100% on all tested variants. A hybrid firewall blocked known patterns but 58 of 570 obfuscated trials bypassed it, and LiDAR poisoning reached 100% against DeepSeek-V4-Flash.

Sources: RIPA preprint