AI Security & Agent Guardrails

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.

Mastercard announced an expansion of its Agent Pay agentic payments program with new trust and intelligence services that give financial institutions and merchants greater context for AI-initiated transactions. The services form the intelligence layer of Mastercard's Agent Pay Trust Framework, which combines identity, intent, controls, execution and intelligence to establish trust across agentic commerce. The first service, rolling out for testing in the U.S., is a probability score indicating the likelihood that a transaction was initiated by an AI agent, which Mastercard said it will strengthen over time with intelligence on behavior, merchant risk, transaction patterns, credential risk and consumer propensity. Mastercard is working with partners including Cloudflare on privacy-preserving environments for understanding AI-driven payment activity, and with Skyfire, a provider of Know Your Agent technology, to help financial institutions and merchants recognize trusted agents. The company cited a projection that one in 10 consumers will routinely use agents to make purchases by 2030, and Ann Johnson, executive vice president of Security Solutions at Mastercard, said the new agentic intelligence and risk insights give people the confidence to say yes however they choose to pay.

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.

IBM outlined how its products will work with Nvidia's new AI-agent safety platform, sending the hybrid-cloud and enterprise-software company's shares up about 0.4% to $221.49 around 10:18 a.m. ET Tuesday. Under the plan, IBM's Agent Identity and HashiCorp Vault connect to Nvidia OpenShell to manage credentials and permissions, while Identity Protection tracks agents across an organization. IBM Fusion adds BlueField-4 hardware and Red Hat OpenShift provides a place to run agents across hybrid-cloud systems, and Nvidia's platform also includes Sentry, which monitors agent behavior and can quarantine an agent that crosses its limits. IBM disclosed no contract value or revenue target for the collaboration, leaving investors to weigh whether customers will pay for the combination at scale. The article first appeared on GuruFocus.

TrendAI, the AI security business unit of Trend Micro Incorporated, announced it has expanded support for Claude's Compliance API in TrendAI Vision One to cover Claude Code and Cowork session transcripts from Claude Enterprise. The move builds on the integration TrendAI announced with Anthropic in June 2026 and adds visibility into how engineering and knowledge-work teams use AI agents day to day. Claude's Compliance API returns session records for Claude Code and Cowork, including prompts, responses and tool calls, each tied to a verified user, and TrendAI now brings those records into TrendAI Vision One so security teams can investigate AI activity in context, correlate it with endpoint, identity, cloud and network signals, and build custom detections. Rachel Jin, Chief Platform and Business Officer and Head of TrendAI, said coding and knowledge-work agents are now doing real work inside the business and security teams should be able to see and investigate that work the same way they do everything else. TrendAI Vision One also continues to ingest activity logs from Claude Enterprise, covering user logins, admin actions and configuration changes, and from Claude Platform, covering admin, system and resource events. The Claude Compliance API integration in TrendAI Vision One is available now for TrendAI Vision One customers using Claude Enterprise.

Nvidia introduced its Open Agent Safety Platform, a new suite that includes OpenShell software and Sentry, a separate hardware-based watchdog designed to monitor autonomous AI agents and quarantine one within milliseconds if it attempts to move beyond its permitted boundaries. The launch drew an unusual challenge from Pope Leo XIV, who told reporters Monday that the contrast between Nvidia's new safeguards and CEO Jensen Huang's opposition to additional government regulation raised concerns that should be taken seriously. Huang has argued that AI safety can be addressed through engineering and market forces rather than new laws, saying earlier this month that the industry does not need additional AI regulation. The debate is increasingly relevant for Nvidia as it expands beyond selling chips deeper into the software and infrastructure used to operate AI agents. Nvidia shares were up about 0.6% in premarket trading Tuesday after closing Monday at $228.86.

Oracle unveiled Fusion Claw, a new agentic governed execution runtime for enterprise AI workloads, alongside 25 new agentic applications that can reason, compute, adapt and execute complex work at scale. Chris Leone, Oracle's EVP of applications development, described the launch on NYSE Live as a leap comparable to moving from an assisted-driving car to a full-self-driving car. The 25 claw-powered applications extend the agentic applications Oracle delivered about nine months ago, taking on more sophisticated, higher-value work such as financial analysis and supply chain operations. Leone cited hospital nurse scheduling that can optimize utilization across five hospitals and trucking supply chain planning that aligns production lines with shipping over an eight-week window. He said governance is the key engineering problem, handled through an enterprise operating envelope that sets risk thresholds, controls, policies and delegation authority, plus an outcome receipt that provides a full chain of evidence for every claw run.

Nvidia has announced an agent safety platform that pairs its open-source OpenShell software fence with a new hardware chip called Sentry, designed to monitor and quarantine AI agents that break out of their sandboxes. OpenShell builds a software boundary around an AI agent, while Sentry sits on the outer perimeter as a hidden guard that the agent cannot detect and can quarantine it the moment it steps outside the fence. Nvidia says the key distinction is that Sentry enforces containment at the hardware level, whereas past escapes, including the Hugging Face incident and attacks tied to Anthropic's Claude, occurred at the application or software level. Nvidia chief executive Jensen Huang confirmed in the report that under the exact same conditions as the Hugging Face incident, which went undetected for months, the platform would have prevented the entire event in milliseconds. The announcement follows Huang's remarks that AI labs calling for a slowdown while simultaneously accelerating AI compute investment are contradicting themselves, and that companies can simply release safe products instead.

President Trump is meeting with several AI leaders at the White House on Tuesday, including Nvidia CEO Jensen Huang and Anthropic CEO Dario Amodei, who will also hold separate meetings with lawmakers on Capitol Hill. The Hill push comes even though the House is in recess until after the midterms and the Senate is about to go into recess, so no imminent action is expected despite legislation already on the table, including a bipartisan bill from Democrat Ted Lieu and Republican Nathan Moran that includes a kill switch among other provisions, and a proposal from Bernie Sanders. Later this week, a Senate subcommittee will hold a hearing titled Rogue AI, securing the homeland against AI agent attacks, and is expected to call several third-party AI researchers who have been identifying these threats. In June, Florida Attorney General James Uthmeier, a Republican, sued OpenAI, and just yesterday sought a temporary injunction requiring the company to stop marketing its products as safe, halt its development, and use third-party guardrails for new models. The discussion noted that litigation lags technology and legislation lags litigation, pointing to social media as an example of how long it took before landmark lawsuits and real legislation emerged.