David Dugan, a longtime advertising executive from Meta, has joined OpenAI as Vice President of Global Advertising Solutions, leading a newly formed ad team. OpenAI stated it will gradually roll out ads within the ChatGPT app over the coming weeks, having already conducted limited tests earlier this year. The initiative has attracted major agencies including Dentsu, Omnicom, and WPP, as well as brands like Wired and Best Buy. ChatGPT boasts approximately 900 million weekly active users. During testing, the cost per thousand impressions (CPM) was around $60, with brands required to commit to ad spend of at least $200,000–$250,000. OpenAI emphasized that ads will not influence conversational outputs and that user chat data will not be sold.

Luma AI has released Uni-1, an image generation model using an autoregressive architecture instead of diffusion-based methods, unifying 'understanding' and 'generation' into a single set of weights. This design enhances complex instruction following, spatial reasoning, and multi-round editing. On the RISEBench benchmark, Uni-1 achieved a total score of 0.51, leading overall and showing stronger performance in logical reasoning subtasks; it scored 46.2 mAP on ODinW-13 object detection. Pricing for generating a 2K image is approximately 9 cents, about 30% lower than competitors. Luma is integrating Uni-1 with its Agents platform for enterprise workflows such as ad production.

Google announced the integration of Gemini AI agents into its Google Threat Intelligence platform for dark web threat intelligence gathering, currently in public preview. The system analyzes approximately 8 to 10 million posts daily, identifying data leaks, initial access broker activities, and insider threats by building customer organization profiles and performing vector comparisons. It can generate high-relevance alerts based on data from the past seven days. Google claims internal testing shows about 98% accuracy, significantly reducing the 80%–90% false positives common with traditional keyword-based tools. Additionally, security operations AI agents are now available in Security Operations, and enterprises can use MCP to build custom security agents under unified governance.

Alibaba International announced the launch of Accio Work, an enterprise-grade AI agent positioned as a 'no-code AI team' for global SMEs, capable of handling long-cycle tasks such as market analysis, design, procurement, store optimization, and inventory monitoring. Officially, Accio has served over 10 million monthly active users since its debut in November 2024 as a B2B procurement engine. Accio Work dynamically orchestrates cross-functional agent teams to work in parallel and provides automated compliance capabilities across more than 100 markets. The platform enables businesses to package workflows into reusable and monetizable skills, emphasizing sandboxing, fine-grained permission controls, and data sovereignty protection. The product is scheduled to launch on Accio.com by the end of March 2026.

Munich-based startup Interloom announced a $16.5 million funding round led by DN Capital, with participation from Bek Ventures and Air Street Capital, aiming to transform 'tacit knowledge' from enterprise processes into a memory layer accessible to AI agents. The company builds 'context graphs' by analyzing operational data such as support emails, service tickets, and call logs, providing AI agents with organization-specific decision-making context to reduce execution errors caused by documentation gaps or staff turnover. Reports indicate deployments at Commerzbank, Volkswagen, and Zurich Insurance; at Commerzbank, knowledge gaps were reduced from 50% to 5%. Funds will support product development and market expansion.

DeepTempo launched and open-sourced Vigil, an AI Security Operations Center project at RSAC 2026, featuring an 'LLM-native' architecture designed to replace closed proprietary AI SOC systems with pluggable multi-agent workflows. Released under the Apache 2.0 license, Vigil offers 13 specialized AI agents, over 30 integration connectors, and more than 7,200 detection rules supporting formats like Sigma, Splunk, Elastic, and KQL. The platform supports incident response, investigation, threat hunting, and forensics, allowing users to extend integrations, rules, and agents via file submissions. It emphasizes transparency and auditability, enabling enterprises to deploy with their own models and existing security infrastructure on-premises.

Qualys introduced Agent Val, an AI agent integrated into its Enterprise TruRisk Management (ETM) platform, designed to safely validate exploitability of vulnerabilities in production environments and drive autonomous remediation cycles. Based on TruConfirm technology, it shifts from CVSS-based hypothetical prioritization to evidence-driven workflows: confirming real attack paths first, then prioritizing fixes based on business context and asset criticality, and revalidating control effectiveness post-mitigation. Qualys claims it reduces remediation noise by over 90% and shortens confirmed vulnerability fix times by approximately 70%, currently covering more than 1,600 CVEs. Deployment requires no additional sensors. This approach aims to shift focus from 'vulnerability count management' to 'exploitable risk management'.

GitHub announced the introduction of AI-powered security detection in pull request workflows to complement CodeQL's static analysis, particularly expanding coverage to ecosystems where semantic analysis has been challenging—such as Shell/Bash, Dockerfile, and Terraform (HCL)—and providing more context-aware fix recommendations. GitHub also revealed that Copilot Autofix repaired over 460,000 security alerts in 2025, with an average fix time of 0.66 hours compared to 1.29 hours without Autofix, representing nearly a 50% improvement in speed. GitHub emphasizes its hybrid detection strategy combining 'traditional rules + AI', aiming to shift risks earlier into the development lifecycle without significantly increasing the burden on security teams.

A research team from North Carolina State University proposed the 'Shallow Safety Alignment Hypothesis' (SSAH), arguing that many LLM safety mechanisms perform only one-time checks early in generation, making them vulnerable to bypass during long responses or multi-turn interactions. The study identifies safety-critical neurons responsible for determining whether a model should respond and proposes freezing these neurons during fine-tuning—preserving the original model’s safety properties while adapting to new tasks, thus reducing the risk of safety degradation due to training. The paper also suggests leveraging redundant units in pre-trained models as an 'alignment budget' to minimize the 'alignment tax'—the performance cost often associated with safety enhancements. The team has made related code and information publicly available and plans to present the work at ICLR 2026.