Tech Logic

Cisco study says multi-turn manipulation can bypass mainstream LLM safety guardrails, involving OpenAI, Anthropic, Google, Amazon and xAI models

Cisco researchers say several mainstream large language models may have their safety guardrails bypassed under multi-turn, iterative conversational manipulation. The disclosed sources collectively point to models such as ChatGPT, Claude, Gemini, Amazon Nova and Grok, but differences in specific model configurations and the extent of impact are only partially mentioned in the provided sources and could not be fully confirmed.

Jun 2, 2026

IEI Announces COMPUTEX 2026 Edge AI Platform Showcase, Focusing on Industrial Automation and Network Resilience

IEI Integration Corp. said it will present “Resilient Edge AI Platforms” at COMPUTEX 2026 at booth P0114 in TaiNEX 2, Taipei Nangang Exhibition Center Hall 2, highlighting the convergence of AI computing, real-time control, and cyber-resilient infrastructure at the industrial edge. Cross-checking across three sources confirms the IEI-related facts; a third source mentions a joint Blaize and Winmate showcase, but it is a different entity and cannot be merged with IEI’s announcement.

Jun 1, 2026

Illinois Passes SB 315 AI Safety Bill: Frontier AI Companies Face Third-Party Audit Requirements

The Illinois legislature has passed SB 315, an AI safety regulatory bill. Its core provisions include requiring frontier AI companies to undergo third-party safety audits, submit safety plans, and report major safety incidents. All three sources confirm the bill has passed the state legislature, and Governor JB Pritzker has said he will sign it. However, the exact scope of the bill, whether it has formally taken effect, and comparative claims such as “strongest” or “first” vary across sources or could not be further confirmed.

May 31, 2026

Hassabis: AGI is still broadly expected around 2030, with 2029 now possible

Google DeepMind CEO Demis Hassabis told Axios after Google I/O 2026 that his view on the timeline for artificial general intelligence (AGI) remains “broadly” around 2030, but he now also sees 2029 as a possibility. Of the three sources, the first two directly corroborate this timeline and its rationale; the third only mentions AI-agent trends at Google I/O 2026 and does not directly address the AGI forecast.

May 29, 2026

Microsoft open-sources Rampart and Clarity: Bringing AI agent security earlier in the development lifecycle

Microsoft released two open-source AI agent security tools, Rampart and Clarity, on May 21, 2026. All three sources confirm that the goal is to shift security checks earlier and support agentic AI security engineering. On specific uses, the sources consistently mention red-team testing and helping developers build safer agents, but phrases such as “assisting incident response” and “handling ongoing breaches” appear in only one source and cannot be confirmed from the provided material.

May 24, 2026

Google I/O 2026 Unveils Gemini Omni: Multimodal Input to Video Generation, with All Three Sources Pointing to a “Video-First” New Model

Google unveiled Gemini Omni at I/O 2026. Three sources consistently confirm that its core capability is to accept multimodal inputs such as text, images, audio, and video, and generate high-quality video content. However, there are differences in how the sources describe its final capability limits, whether it already supports the broader idea of “anything from any input,” and how it relates to existing video generators. Some details cannot be confirmed from the provided sources.

May 20, 2026