AI Security & Agent Guardrails

Zscaler CEO Jay Chaudhry said AI agents will become the weakest link in corporate security, warning that agents running loose on corporate networks can be hijacked and operate at machine speed with no breaks. In a Yahoo Finance interview, Chaudhry said more than 50% of Fortune 500 companies are Zscaler customers and that the company is extending its zero-trust architecture to agents, giving them access only to specific applications and services. He said enterprises are not ready for tools like Meta's Muse, though his team has been testing it, and he rejected OpenAI CEO Sam Altman's suggestion that the world should accept some bad outcomes for AI's benefits. Chaudhry also said Zscaler's investor day, its first since 2021, will share visibility into the AI and agentic contribution to its financials, after JPMorgan asked for greater clarity on timing and magnitude. He added that going public makes companies more responsible and that Zscaler has been public for eight years.

Promevo launched Insights by Promevo, a platform it calls the first to unify AI agent governance, Gemini Enterprise adoption and Google Cloud cost management into a single view. The platform delivers real-time observability across three areas: AI license optimization that identifies idle or underutilized licenses and quantifies reclaimable spend, AI agent governance and security that inventories custom-built agents and maps their permissions and data access, and AI cost attribution that connects every agent to its underlying Google Cloud resource consumption and costs. Promevo says native consoles offer fragmented data with limited 28-day lookback windows, while Insights consolidates these silos with extended historical retention, and it does so without ever reading or storing an organization's sensitive prompt or response content. The same view extends beyond Gemini Enterprise into the rest of a customer's Google Cloud environment, covering cloud FinOps, asset discovery and hierarchy, and cross-project IAM access intelligence. CEO Karthik Kripapuri said customer organizations already run thousands of agents, and the platform is available today for customers using Google Workspace and Gemini Enterprise, with expansion planned in the coming days and weeks.

Mastercard announced a probability score for AI-initiated transactions on September 30, positioning the capability inside its Agentic Commerce Trust Framework alongside Cloudflare for web and payment signal feeds and Skyfire for Know Your Agent verification. The probability score adds a second layer to the trust stack that agent commerce requires, evaluating each AI-initiated transaction dynamically to produce a trust score at the moment of execution rather than asking whether a transaction is legitimate. It builds on Skyfire's Know Your Agent framework, which Mastercard integrated earlier this year and which issues credentials and verifies agent identity at the point of transaction, a foundation also pursued by Baselayer, which raised $35 million to build Know Your Agent infrastructure. With the probability score in place, the trust infrastructure for agent commerce now has four visible layers: Know Your Agent identity verification, transaction-level probability scoring, payment execution with guardrails such as Stripe's Agentic Commerce Suite and Visa's Intelligent Commerce, and consumer protection signals that remain nascent and fragmented.

The Electronic Transactions Development Agency, or ETDA, has announced its artificial intelligence governance direction for 2027, elevating its AI Governance Practice Center, or AIGPC, into a central mechanism to drive an AI Governance Sandbox, a space where organizations can test AI solutions and policy approaches against real use cases in a controlled environment in order to uncover risks, set assessment criteria, and refine systems before wider deployment. Ms. Rodjana Lamlert, an ETDA advisor and director of the AIGPC, said the key challenge is not making everyone aware of AI or getting them to adopt it, but rather that organizations differ greatly in their levels of knowledge, readiness, and ability to manage AI-related risks. This is summed up as G-A-P, covering three gaps: the gap in understanding of AI governance, readiness to apply AI, and the translation of ethical principles into actual practice. The core idea is not to wait for problems to arise and then fix them, but to identify risks before AI is actually deployed. ETDA also plans to work with Singapore to develop a Thai-language test prompt set designed to understand Thai language, culture, and social context, and to push for an AI Incident Exercise to rehearse responses to cross-border AI incidents, such as AI scammers who fake images, voices, or messages to impersonate individuals or government agencies. In 2026, ETDA laid the groundwork for scaling up its efforts by working with a network of more than 140 agencies spanning 20 ministries, developing an AI Governance Guideline, a Toolkit, and an AI EIA Playbook together with the banking sector to test an Ethical LLM, and it aims to develop AI governance skills in more than 120,000 personnel.

Nvidia Corporation is adding the Open Agent Safety Platform to its AI infrastructure stack, an open software and reference system for securing AI agents from testing to deployment that combines OpenShell software with Nvidia Sentry on BlueField DPUs to monitor, control, and isolate potentially harmful agent activity. The launch follows recent incidents involving AI agents from OpenAI and Anthropic, including an OpenAI agent breaching Hugging Face, and Nvidia said the new platform could have prevented that breach by controlling agent permissions and isolating suspicious behavior. More than 100 organizations, including major technology and software enterprise companies, are involved in the platform. Nvidia said the main uncertainty for investors is how it will monetize the platform, since OpenShell is open source and works across different CPU architectures, while Sentry is more directly tied to the company's BlueField hardware, and it is still too early to view the platform as a meaningful short-term earnings driver. Nvidia's forward GAAP P/E of 22.79x sits about 56% below its 5-year average of 51.72x, and its forward price-to-sales ratio of 13.33x is about 32% below its 5-year average of 19.70x. As of July 26, Nvidia held $22.44 billion in cash and cash equivalents and $34.14 billion in marketable debt securities, against about $33.37 billion in total debt, while hedge fund ownership rose from 275 funds at the end of Q1 2026 to 285 funds at the end of Q2 2026 and short interest stood at just 1.27% of float as of September 15, 2026.

Apple said on the 2nd that it will modify the operating software of its Mac personal computers so that users can more clearly recognize and respond when artificial intelligence agents request access to all data on a Mac. On its website, Apple pointed out that some developers are using the "Full Disk Access" feature in ways that could put users at risk, and stated that it will introduce additional controls going forward. The company said that "as AI agents become more capable and autonomous, the risks associated with such broad access increase substantially," and that it is working to ensure users can clearly understand these risks before granting access, so they can make well-informed decisions about their data and privacy. The Mac is designed to be more flexible than the "sandboxing" used on the iPhone and iPad, allowing apps such as cloud backup services to access all data on the device with the user's permission. As for Meta Platforms' AI agent "Muse," some users have criticized it for accessing highly sensitive personal information.

Nvidia added $150 billion to its share repurchase authorization, the largest increase of its kind in history, leaving $235 billion of buybacks still waiting to be executed. The company says it expects to work through the full $235 billion by the end of fiscal year 2028, a window that overlaps with its guidance for 70% revenue growth in fiscal 2028. Nvidia also unveiled the Open Agent Safety Platform, built from OpenShell, open source software that draws a runtime boundary around autonomous AI agents and runs on Nvidia's Vera CPUs, and Sentry, a reference design on BlueField-4 DPUs that quarantines straying agents in milliseconds. Anthropic, SpaceXAI, Scale AI, Salesforce and SAP have signed on, and Nvidia counts over 100 organizations working with the technology. Hedge fund ownership climbed to 285 funds from 275 in the prior quarter, while short interest sits at just 1.27% of the float and the shares trade at 24.88 times forward earnings.

Meta Platforms joined other large AI developers at the White House to sign a voluntary AI safety accord. The agreement commits participants to independent testing of AI systems, stronger internal controls and clearer public transparency measures, and Meta also agreed to board-level oversight of key AI risks, aligning governance of its AI work with the accord's safety standards. The voluntary accord points to areas that can drive significant costs for Meta, such as independent audits, tighter internal controls and board-level AI risk oversight, which can increase compliance spending and slow the rollout of some systems while potentially reducing the chance of large fines or forced product changes later. Meta is trying to sell Muse-powered tools and the Meta Enterprise Platform into businesses that care about compliance, security and data handling, and clearer safety governance can support that effort by giving CIOs more confidence in how Meta builds and tests AI. The next real test will be how Meta describes AI risk oversight and testing in upcoming SEC filings and earnings calls, especially after appointing a Chief Enterprise Platform Officer and launching Meta Enterprise Platform on 28 September 2026.

OpenAI, Google and Meta have joined an agreement that provides for third-party audits of the safety of artificial intelligence systems, marking a significant collaboration among major technology companies to allow independent bodies to assess the safety of AI.

Nvidia has launched the Open Agent Safety Platform, a security platform for AI agents that it describes as akin to a browser for agents, acting as an intermediary that controls permissions so agents can access only the resources and systems necessary for their tasks. The architecture consists of two main parts: Nvidia OpenShell, a processor-level control system that defines permission boundaries and restricts behavior to prevent AI from intruding into unauthorized systems, and Nvidia Sentry, a monitoring system that runs separately on network chips to detect and immediately suppress abnormal behavior without drawing on primary computing resources. The platform is structured as a Reference Design emphasizing open-source software, giving global partners such as Microsoft, Cisco, Oracle, Dell, Intel and Anthropic the opportunity to build on it and develop their own AI security products. The launch aligns with the view of Jensen Huang, CEO of Nvidia, who sees AI safety concerns as fundamentally an engineering problem that can be solved through effective system governance design, rather than by halting innovation.

Google debuted its Gemini 4 Argon frontier AI model on Wednesday, saying it delivers high-performance capabilities for tasks including software engineering, knowledge work, and cybersecurity. The company is initially releasing Gemini 4 Argon to trusted members of its Fairwind Program, which allows them to use Argon to find flaws in their software and fix them before Google makes the model more widely available, and it is also working with the US government's voluntary pre-release program for frontier models. Google is jumping into the AI pricing wars as well, offering Gemini 4 Argon at introductory rates of $2 per million input tokens and $10 per million output tokens, well below Anthropic's highest-end Fable 5.1 at $10 per million input tokens and $50 per million output tokens and its Opus 5.5 at $4 per million inputs and $20 per million outputs, and below OpenAI's top-tier GPT-6 Astra at $10 per input and $50 per output, though OpenAI's new GPT-6.1 Sol matches Google at $2 per input and $10 per output. Google says the software already powers internal workflows ranging from coding to research and writing quality, and that it plans to release the software in phases before it is available for enterprise and customers, with Gemini 4 Argon eventually underpinning the majority of Google's services, similar to how the company incorporated Gemini 3 across platforms ranging from Search to YouTube. Ahead of the release, Bloomberg reported that some within Google worry Gemini 4 Argon isn't as powerful as Anthropic's or OpenAI's offerings, while others say it is; Alphabet stock was flat on the news.

Nvidia introduced the Open Agent Safety Platform, adding OpenShell and Sentry tools for autonomous AI oversight. The safety suite launches with partners including Bedrock Data and TrendAI to help enterprises set and enforce AI usage boundaries. Nvidia CEO Jensen Huang joined a White House event where leading AI developers signed a voluntary accord on safety standards. The platform pairs agent runtime controls like OpenShell with hardware level oversight from Sentry on BlueField 4 DPUs, extending Nvidia's full stack pitch beyond GPUs, CPUs and networking into governance and compute for agentic AI projects. Nvidia carries a market value of about $5.5 trillion.

Dr. Kobsak Pootrakool, Vice Chairman of Bangkok Bank, or BBL, said in a personal Facebook post that the race for leadership in artificial intelligence between the United States and China is intensifying, with the US confident it still leads China by roughly 6 to 12 months and ready to accelerate technology development to preserve that lead, a factor that will be key in shaping the direction of the global economy and in gaining an edge in geopolitical competition. President Donald Trump invited leading private-sector figures in AI to discuss ways to maintain US leadership alongside building confidence in AI safety and developing infrastructure, especially data centers, which must win acceptance from local communities. Dr. Kobsak said that although the US has funding to support AI development, it faces pressure from safety concerns and community impact, as well as opposition to data center construction in many areas. As a result, the White House Accord on Super Intelligence was drawn up under a Joint Commitment on Frontier Responsibilities, joined by leading companies including Google, Anthropic, Meta, OpenAI, xAI and Nvidia. It is a voluntary safety agreement, which President Trump described as a moral commitment by AI developers that could later be developed into law or regulation. There are also guidelines setting out a four-layer AI safety review system: controlling and monitoring model operations within organizations, setting up internal teams to review and fix flaws, assessment by independent external auditors, and oversight by independent company boards. At the same time, the US is supporting the expansion of data centers by setting guidelines for project developers to deliver benefits to communities, in areas such as education, quality of life, financial support and reducing the tax burden on local residents. President Trump also favors having government agencies rename AI as Super Intelligence, or SI, and plans to set up an AI Board, a committee of about 10 people to oversee and accelerate AI development, as well as to announce a new person in charge of AI policy. Details of the names and their powers still need to be followed. Dr. Kobsak views that the race for AI supremacy is entering a crucial phase, in which the ability to manage safety risks and social impact will be a key condition for accelerating technology development, which could affect the global economy and people's lives going forward.

President Donald Trump signed a voluntary AI safety agreement with the leaders of several major AI companies at the White House on Tuesday, calling the pact "morally binding." The White House Accord on Super Intelligence, also described as a "Joint Commitment on Frontier Responsibilities," was signed by Trump along with Alphabet CEO Sundar Pichai, Anthropic CEO Dario Amodei, Meta Platforms CEO Mark Zuckerberg, OpenAI president Greg Brockman, SpaceX and xAI CEO Elon Musk, and Nvidia CEO Jensen Huang. The accord sets out four layers of controls for companies developing frontier AI models: internal systems to monitor AI capabilities and alignment during training and deployment, particularly for cybersecurity, biosecurity and chemical risks; independent external auditors or evaluators to assess whether safeguards work as intended; a separate independent committee of each company's board to oversee the process and receive reports from internal teams and outside auditors; and internal teams responsible for keeping monitoring and detection systems functioning. The agreement leaves open the possibility of future regulation, stating that "it may make sense to codify these steps into laws or regulations" over time, and the participating companies agreed to meet regularly to develop safety standards and best practices. Zuckerberg called the agreement a "significant positive step" on X, saying multiple layers of audits and reviews could give the public more confidence in advanced AI systems.

U.S. President Trump on the 29th publicly posted on social media a document he signed with the heads of artificial intelligence companies. The AI companies stated they will "introduce robust internal governance systems for monitoring models," and pledged to carry this out, saying that "implementing controls and audits is important for securing a safe future for everyone." They said they will hold regular meetings to improve safety.

Aembit announced support for Okta's Cross App Access protocol to manage enterprise AI agent access, and revealed that Helen Riley, an executive at Alphabet's X, is joining its board of directors. The XAA integration is intended to streamline authorization and oversight for automated agent workflows used by corporate customers. The XAA integration and Helen Riley's board role are only part of the broader story around Okta's identity platform, and Simply Wall St flagged one warning sign for Okta. Okta positions itself as an identity partner for enterprises that want a single control point governing how humans and software agents reach critical systems, so moves around Cross App Access plug directly into how the firm aims to sit between corporate users, AI tools, and cloud infrastructure. The article points toward a $122 fair value for Okta.

Nvidia is rolling out a new AI safety software platform that uses an independent third-party model watchdog to observe the reasoning of AI agents, according to Creative Strategies CEO and Principal Analyst Ben Bajarin. Bajarin said the platform's core function is to have a separate model monitor an agent's reasoning in real time and flag when it starts to go off the rails or reasons toward actions it should not take. The system can be configured with enterprise-specific permissions and policies to enforce guard rails, an approach Bajarin called fairly novel because it inspects reasoning rather than only judging outcomes. The software news came alongside Nvidia's move to boost its buyback by $150 billion. Bajarin said AI safety concerns voiced by OpenAI and Anthropic are fair, warning that many enterprise agents are currently running ungoverned and that better governance and security are needed to defend against frontier hacks.

Nvidia introduced its Open Agent Safety Platform on September 28, 2026, a framework that shifts AI security from policy-based promises to silicon-level enforcement using the OpenShell open-source runtime and Sentry, a BlueField-4 DPU watchdog, with more than 100 partner organizations already involved. The initiative faces a strict liability stance from the Federal Trade Commission, where Chair Andrew Ferguson said at the Reuters Momentum AI conference in Austin on September 25 that the man who wielded the hammer ought to suffer the consequences of his conduct, rejecting rogue agent defenses. The tension is sharpened by the Connecticut AI Responsibility and Transparency Act, which begins enforcement on October 1, 2026 and holds deployers liable for AI outcomes even when they use third-party vendor tools, with no state guidance yet on whether voluntary frameworks like OASP satisfy those obligations. Insurers have pulled back on agent-related coverage through Verisk AI exclusion endorsements, and only 22% of enterprise contracts currently offer uncapped indemnity, leaving the liability burden with deployers regardless of hardware safeguards. Market participants should watch whether the Connecticut Attorney General issues guidance on how voluntary frameworks interact with CAIA compliance, and whether new insurance products emerge to bridge technical safety and legal liability.