Infostealer malware is being used to replay stolen Claude session cookies into paid accounts without triggering two-factor authentication or Single Sign-On (SSO) protections. These affected accounts are self-serve with direct card billing, outside the governance of corporate identity providers, making it impossible for IT admins to revoke sessions. Anthropic has responded by notifying users, signing out stolen sessions, removing saved payment methods, and refunding fraudulent charges. However, the sessions can potentially access Gmail inboxes, conversation histories, files, and connected services authorized by users — posing a significant security risk for corporate data. Attackers often obtain these cookies from infected devices, sometimes via pirated software or spoofed download sites. Experts recommend treating AI service sessions on compromised machines as breached, revoking any tenant permissions possible, and moving heavy users to managed organizational accounts with tighter controls like OAuth grant restrictions and device-bound session credentials. Endpoint security combined with identity governance and AI policy enforcement is critical to closing this security gap.
New survey data from VentureBeat’s July VB Pulse reveals that 39.4% of enterprise buyers plan to evaluate AI accelerators from providers other than Nvidia—such as AWS Trainium, Google TPU, AMD Instinct, Intel Gaudi, or custom ASICs—within the next year, surpassing the 25.3% considering Nvidia's upcoming Blackwell GPUs by 14 points. While Nvidia remains dominant in production, organizations are diversifying their AI accelerator strategies and increasing investment in existing infrastructure rather than rushing platform changes. Enterprises are also intensifying usage of platforms like Microsoft Azure, Google’s Gemini, and OpenAI, with improvements in operational reliability and throughput. Interest in non-Nvidia chips is particularly strong among strategic buyers and smaller companies, reflecting a shift to multi-vendor approaches and higher expectations for infrastructure performance and cost efficiency. Additionally, the market is seeing growth in open-source AI production systems and specialized neocloud providers, which offer new strategic options. Overall, enterprises are becoming more capable AI infrastructure operators, prioritizing reliability and flexibility, and preparing for future platform decisions from a position of greater maturity and optionality.
OpenClaw 2.0 marks a significant evolution from a personal AI coding assistant into a collaborative platform designed for enterprise use. This update introduces a new browser-based interface that unifies conversations, files, approvals, and agent activity into a shared workspace. Key features include multi-user cloud sessions, role-based permissions, sandboxing, and enhanced security controls, enabling persistent and supervised agent workspaces for teams rather than individuals. The platform now supports a broader employee base with a conversational UI resembling popular AI chat services, while maintaining transparency into the underlying processes. With collaborative session ownership, detailed audit trails, and integration with external credential stores, OpenClaw 2.0 aims to serve as a secure, operational control plane for organizations managing AI agents. Its security posture is configurable, requiring deliberate policy to achieve enterprise readiness, contrasting with competing projects like NanoClaw that embed stronger isolation by default. Despite being developed independently under the OpenClaw Foundation, OpenClaw 2.0 leverages backing from major tech partners and positions itself as a shared digital workspace for AI-enhanced productivity across enterprise teams.
Perplexity has unveiled hybrid compute for its AI platform, Computer, enabling a single AI agent to split tasks between cloud-based frontier models and smaller models running locally on Apple silicon Macs. This setup routes sensitive data to local devices, ensuring it never leaves the user's machine. The system uses a Privacy Gate, a classifier running on the device that detects personally identifiable information to prevent sensitive data transmission to the cloud. Users can choose whether flagged data is processed locally or shared. Targeted at enterprises, this approach balances cutting-edge AI intelligence with privacy and security, especially for confidential tasks like legal work, financial modeling, and marketing analytics. The platform supports models from Google, Alibaba, and a proprietary Perplexity-trained version, emphasizing local inference to mitigate geopolitical risks. Perplexity also offers robust controls for enterprise compliance and transparency, including audit logs and opt-out data policies. This innovation addresses the critical challenge of data privacy in AI adoption, blending cloud power with local security to meet professional standards.
Egiziago Cioffi, CEO of SynSphere Italia, faced a critical problem with his Azure OpenAI assistant: it returned sensitive SharePoint content to users without proper permission checks. Even though the assistant passed all quality evaluations, a low-privilege user could still access data they shouldn’t see because the retrieval pipeline used the indexer's permissions, not the requester’s. This issue is common in many AI retrieval deployments where native document-level access controls are missing or incomplete. Azure AI Search now includes built-in ACL trimming for document access, but it isn’t universally applied, especially in custom pipelines like Cioffi’s. To fix this, Cioffi implemented a simple but effective query-time filter that checks user permissions before retrieval, narrowing the assistant’s scope without the need for a new identity platform. This filter allowed the assistant to continue resolving about 60% of inbound emails while enforcing proper access boundaries. The case highlights a blind spot in AI assistant evaluations—they usually check for correct answers but not whether responses respect access controls. The lesson is clear: AI retrieval systems must verify whose permissions they enforce at query time to prevent unauthorized data exposure. A straightforward two-account test comparing access levels can reveal this critical gap quickly and should become a standard security check before deploying any AI assistant.
Enterprise AI is advancing rapidly, transitioning from simple assistants to autonomous agents capable of complex reasoning, tool invocation, and multi-step workflow execution with minimal human input. Unlike traditional software that follows set logic, AI agents dynamically decide how to reach goals, making security more challenging. Traditional authentication confirms identity but does not guarantee ongoing trustworthy behavior during AI operations. This leads to unique risks such as goal drift, excessive tool use, memory poisoning, and manipulated context, which can compromise workflows and enterprise systems.
To address these, enterprises must adopt 'runtime trust' — a continuous monitoring and validation approach ensuring AI agents' actions align with organizational policies throughout their operation. Key components include intent validation, behavioral monitoring, strict policy enforcement, least-privilege access, and human oversight for critical decisions. This approach extends to securing interconnected AI ecosystems, including knowledge bases and external AI services, with strong visibility via runtime logging and audit trails.
Organizations can start by auditing AI capabilities, enforcing dynamic permissions, monitoring anomalies, and integrating runtime trust with existing security frameworks to safely harness autonomous AI's benefits. As AI systems grow more independent and collaborative, continuous trust verification will be essential for mitigating risks and enabling responsible deployment of enterprise AI.
Over the past two years, the way software engineers work has fundamentally shifted. Writing code from scratch is no longer the main challenge. With smart AI tools integrated into development environments, engineers can quickly generate initial implementations for complex systems like streaming pipelines or API integrations.
AI agents can explore codebases, write tests, debug, and suggest improvements, freeing engineers from routine coding tasks. Instead, their role evolves to guiding and overseeing the AI’s output.
Think of AI agents as engines that need clear intent to perform meaningful work. Left unchecked, these agents accumulate errors and assumptions, causing their output to drift away from correctness. Human intervention — through clear feedback like test failures or precise contracts — realigns their work.
Enterprise software complexity resembles the "three-body problem" in physics: multiple interacting systems create unpredictable behaviors. Changes in one system can ripple unpredictably through others.
Because of this complexity, engineers focus on creating strict boundaries and rules — such as semantic layers, immutable logs, and data contracts — that keep AI-generated code reliable. These boundaries transform chaotic problems into manageable domains.
Ultimately, as AI tools generate more code, engineers’ true value lies in designing these containment frameworks and ensuring generated solutions are trustworthy and aligned with business needs. They become architects of equilibrium, maintaining order amid accelerating complexity.
For the past two years, the prevailing belief in enterprise AI was that greater autonomy meant better performance. The idea was to build agents that could independently plan, decide, and act across multi-step workflows with minimal restrictions. However, in real-world production settings, this approach is proving problematic. Successful companies are now focusing on creating AI agents with clearly defined responsibilities that operate within strict rules.
Recent forecasts reveal that over 40% of agentic AI projects may fail by 2028, not due to AI capability but because of rising costs, unclear business value, and insufficient risk controls. Governance maturity in responsible AI remains low, with only about 30% of organizations achieving advanced control measures. This gap between AI capability and governance is reshaping the competitive landscape: the priority has shifted from deploying the most autonomous agent to building trustworthy systems that satisfy risk, compliance, and legal teams.
Full autonomy often breaks down in production because autonomous decisions are hard to trace and audit, especially in regulated environments like finance or healthcare. Integration complexity arises when legacy workflows must be rebuilt to accommodate AI agents acting without human input. Enterprises that dive into this without a comprehensive governance strategy tend to stall or cancel projects.
Leading enterprises adopt four key governance patterns: narrow-scope agents instead of broad ones, human checkpoints before critical decisions, built-in decision traceability, and active data sovereignty to limit risk exposure. This approach balances autonomy with accountability and reduces the risk of costly errors or compliance breaches.
A practical framework for evaluating AI agents involves asking whether action decisions can be reconstructed, agents have bounded responsibilities, checkpoints exist before decisions execute, and data access is properly contained. This framework helps businesses scale AI with calibrated control rather than unchecked autonomy.
Ultimately, the companies that win with agentic AI by 2027 will be those that earn the trust of legal and compliance teams through disciplined governance embedded from the start. It’s not about maximum autonomy but smart orchestration of agents with clear oversight and accountability.
Slack, owned by Salesforce, has launched Slack Code, a new feature that integrates AI coding agents like Anthropic's Claude Code, Cognition's Devin, GitHub Copilot, and Vercel's agent directly into dedicated Slack channels. This allows entire teams to collaboratively watch, guide, review, and ship software, moving AI coding out of individual terminals into a transparent, multiplayer workflow. Slack Code creates project-specific channels where all work, including code diffs and live previews, is visible and archived for audit. The approach shifts the bottleneck from coding itself to judgment and creativity, expanding participation across roles such as engineers, product managers, and designers. Slack emphasizes security by restricting agent permissions to those of the invoking user, ensuring no cross-team data leaks. This launch positions Slack as a key platform for collaborative AI-driven development amid Salesforce's strategic pivot toward AI-powered enterprise solutions. Leaders in the space see a future where multiplayer channels and single-user terminals coexist, with a growing preference for collaborative workflows fostering higher quality and efficiency in software development.
TrueFoundry, a San Francisco-based B2B machine learning startup founded by former Meta engineers, has unveiled TrueForge, an open source AI agent harness licensed under MIT. This tool enables greater developer control and cost efficiency in enterprise AI deployments by supporting any preferred AI model and allowing integration into commercial products. Benchmarks on DevRev’s Enterprise-Bench demonstrate TrueForge's ability to complete multi-step tasks at 30% to 75% lower cost compared to Anthropic’s Claude Managed Agents, using the Claude Opus 4.8 model or the open source GLM-5.2 LLM. TrueForge supports flexible deployment from local developer machines to shared enterprise installations and emphasizes efficient context management techniques to reduce compute waste and expenses. While the harness itself is free, enterprises can add TrueFoundry’s commercial AI Gateway for centralized governance, identity, and access controls. TrueFoundry aims for TrueForge to serve alongside other harnesses in a vendor-neutral, cost-effective manner. Existing customers like NetApp and Automattic have benefited from the solution. With $21 million in funding and acquisitions like Seldon AI, TrueFoundry continues growing its presence in enterprise AI infrastructure, blending traditional ML, LLMs, and agent orchestration with governance and monitoring capabilities.
