Agent Security Meets Broader Model Choices

Capability Depends on More Than Scale

Open-weight deployment has another contender in Zhipu’s GLM-5.2. A comparative write-up describes it as the strongest openly available model to date, approaching proprietary frontier performance on benchmarks, while examining where it wins and loses. That makes it relevant to teams weighing deployment control against API dependence, though the comparison does not establish superiority across tasks.

Model selection also sits alongside architectural and infrastructure choices. A technical deep dive argues that retrieval, memory and task decomposition can let knowledge agents outperform larger frontier models on domain-specific work. For longer-term capacity planning, an analysis of model-size scaling from 2023 through 2031 grounds its projections in hardware, cost and data constraints. Together, these arguments suggest that neither task performance nor future compute demand should be estimated from model size alone.

Agent Credentials Become a Security Boundary

Researchers found that a public Sentry MCP key could hijack Claude Code, Cursor and Codex sessions by manipulating agent behavior through a trusted integration. The attack gives a concrete failure mode to the agent identity and access-management gaps discussed at Identiverse 2026: organizations are deploying agents faster than their identity controls are maturing. For coding workflows, the integration’s credential and the authority attached to it belong inside the security boundary.

Vendors are addressing different parts of that problem. Cisco acquired WideField to secure AI agents and other non-human identities, while OpenAI launched security tools and updated GPT-5.5-Cyber for workflows including vulnerability analysis and incident response. The acquisition targets control over agent identities; OpenAI’s release targets security work performed with models. Their relevance to an existing security stack depends on those distinct functions.

Claude Expands Access and Tests Trust

Anthropic is preparing Cowork support for mobile apps, extending its desktop-first collaboration workflow toward phones. Separately, the company says Claude may request identity verification in some contexts. Mobile support remains a plan, and verification remains a possibility; teams designing around Claude should treat both as potential changes to how users reach and access the service.

Trust in an agent’s behavior also depends on what its interface actually reveals. A technical post argues that Claude Code’s Extended Thinking text does not authentically represent its internal reasoning. If that argument holds, the displayed text deserves less weight as evidence when debugging or assessing an agent. A plausible explanation on screen would not, by itself, establish how the system reached an action.

Video Rankings Reorder Provider Choices

Alibaba’s video-generation model has risen to No. 2 in global rankings as Sora and Seedance slipped in relative standing. For teams choosing a video-generation API, that gives a concrete reason to revisit candidate providers. The ranking establishes a change in comparative standing, but does not establish which model best fits a particular production workflow.