AI & Automation Briefing - August 5, 2026
Nvidia-Led Security Alliance for AI Agents Publishes First Proposals One Week After Launch
The Open Secure AI Alliance (OSAA), a group spearheaded by Nvidia and now exceeding 120 member companies, has already produced its first formal proposals through a working group called the Shared AI Findings Exchange (SAFE), managed by the Linux Foundation. The initial proposals, presented at Black Hat in Las Vegas, cover confidential reporting of AI cybersecurity incidents, notifying affected parties, and conducting blame-free post-mortems. Members are also cataloging open source tools relevant to AI agent security. Contributions on the table include Nvidia's LLM vulnerability scanner Garak, Okta's agent identity work, Red Hat's agent governance efforts, Amazon's Strands Agents builder and Cedar authorization language, and Hugging Face's participation following a recent incident involving a rogue OpenAI model on its platform. The alliance includes Adobe, BlackRock, Cisco, Intel, Microsoft, and Visa, but Anthropic, OpenAI, and Google are absent as members despite OpenAI and Google signing the original open letter that led to the group's formation. For builders of agentic workflows, the OSAA's direction signals that security standards for AI agents are moving toward formalization, which will affect how multi-agent pipelines and integrations must be designed.
PAST-Bench Tests Whether AI Agents Actually Improve From Retained Experience
Researchers from Princeton published PAST-Bench, a benchmark designed to measure whether personal AI agents genuinely improve across sessions by reusing stored memory, tool routines, and learned skills. The benchmark covers 26 scenarios and 204 episodes, running agents through sequential tasks with retained experience enabled versus disabled to isolate the effect. Testing seven base models across four agent frameworks, the researchers found improvement is real but inconsistent across capability types. Agents showing the same overall performance gain can differ significantly in whether that gain actually traces back to the intended save, retrieve, and update process. The team also released Hermes+, an extended agent framework built on those findings, which applies five targeted interventions across the agent loop and shows the strongest gains on tasks requiring outdated stored state to be replaced. Results remain model and capability dependent.
LLMs Consistently Fail to Delete Code, Creating Hidden Technical Debt in Automated Workflows
A new benchmark called CanItDelete tested leading LLMs on 200 real-world tasks where deletion was the only required edit. Even top performers struggled. Claude Opus failed 21% of tasks, GPT-4o Sol failed 26%, and several others failed roughly one in three. A common failure pattern, which the researchers call Guard-and-Go, sees models keep obsolete logic active behind a conditional guard rather than removing it outright. Across five leading SWE-bench Verified submissions, 29% of passing patches used this approach, meaning they passed automated tests but were not actually merge-ready. When researchers required the deleted code to be fully absent, resolution rates dropped 21.3 points and a third of previously accepted patches failed. On the training side, adding deletion-focused examples at just 0.7% of training data cut incomplete deletion errors by 13.9 points and improved SWE-bench Verified scores by 5.3 points. For anyone using LLMs to generate or maintain workflow scripts, this research confirms that AI-generated code requires explicit review for what was supposed to be removed, not just what was added.