AI Risks Spread From Outages to 401(k)s as Businesses Weigh Safety and Adoption

StackGen found AI agents have deleted data, databases or live systems autonomously. Those agents acted with valid credentials, meaning traditional monitoring did not identify anything unusual until the damage was done. AI is embedded in claims processing, coding, customer support, decision support, fraud detection, HR, risk analysis and supply chain planning, increasing the chance of outages and unintended outcomes. Attackers can use AI to scale deepfakes, phishing, social engineering, reconnaissance and exploit development. AI systems may also change behavior over time, creating drift and explainability gaps when data sources, integrations, models and prompts change. Employees and teams may use unapproved AI tools with sensitive information or in business-critical workflows, creating shadow AI risk.

The TechRadar report said AI adoption is creating operating dependencies faster than governance is maturing. If an AI-enabled workflow fails, produces incorrect decisions or becomes unavailable, organizations may not know the business impact or have a manual fallback. Companies need to understand what depends on AI, what happens when those dependencies fail, what the business stands to lose, and where action matters most. They also need to determine whether backup models can be used during disruptions and put deterministic or manual solutions in place as workarounds. If AI is the only option, it may be a single point of failure.

The report outlined four business questions: what is impacted if an AI-enabled process fails or produces an incorrect result; what happens next as the failure propagates; what is the financial exposure from disruption, error or delay; and what should be prioritized for additional controls, human oversight or fallback processes. It said enterprises must decide which processes can tolerate answers that are "probably right" and which cannot. An AI-generated recommendation used to inform a decision may tolerate uncertainty; a process that executes a financial transaction, determines a regulatory obligation or controls a critical operation cannot. Explainability should also be a buying criterion. Harvard Business Review says enterprises own the risk even if they outsource AI technology, pointing to lawsuits against Cigna, iTutorGroup, Peloton and Workday as examples of courts and regulators holding users responsible when tools discriminate, mishandle data or harm customers.

For retirement savers, CNBC reported that the AI trade has been volatile as investors weigh questions about the pace of AI development, capacity constraints and safety concerns. John Sedunov, a professor of finance at Villanova University, said a slowdown in AI innovation, adaptation or use could have downstream implications for companies in the average 401(k) investor's portfolio. Five technology giants—Nvidia, Apple, Microsoft, Alphabet and Amazon—accounted for about 30% of the S&P 500 as of Wednesday's market close, according to Morningstar. That means a relatively small group of mega-cap companies can have an outsized effect on stock returns for workers invested in an S&P 500 index fund. Zachary Evens, a manager research analyst for Morningstar, said a slowdown would represent a pullback in the AI stocks or large-cap technology stocks that make up a large portion of investors' portfolios today because of concentration within S&P 500 index funds.

AI exposure is not limited to those household names. It can show up in companies that supply chips, power, data centers and other infrastructure for the AI buildout, including industrial and smaller companies. Marta Norton, chief investment strategist at Empower, said the AI infrastructure build has so many supply chain elements that investors can look at industrials or small-cap space and still have AI exposure. Target-date funds, the default option in many 401(k) plans, can also include the S&P 500 and other funds with AI-linked companies. Norton said target-date investors' asset allocation varies over time, and as they shift from equity to fixed income, some AI exposure will come down naturally, but much of the growth exposure remains attached to AI because the U.S. market is attached to AI. Nicolas Abrams, a certified financial planner and CEO of investment advisory firm Opulentia, said "Nine times out of 10, all of your money is not at risk with AI" and that even with volatility, investors have a whole portfolio of other investments. CNBC said the appropriate response is generally not a sudden change, but reviewing top holdings, rebalancing if overweight and considering a portfolio-bucketing approach for near-term spending needs.

At Salesforce's annual Dreamforce conference in San Francisco this week, CNBC reported that keynote conversations between CEO Marc Benioff and the heads of Anthropic, OpenAI and Nvidia delved into the AI safety debate. Nvidia CEO Jensen Huang, on stage with Benioff, urged frontier labs to "run as fast as you can." But attendees said they are having trouble taking advantage of existing technology. Alec Bronston, a senior Salesforce director at Chicago-based retail data company Spins, said "It's already hard enough to keep up" and that a slowdown would give companies "a lot of opportunity to just even catch up and get our feet wet." Days before the conference, an Anthropic researcher resigned and said top labs were "gambling with our lives," leading Anthropic's Dario Amodei and OpenAI's Sam Altman to propose safety initiatives and push for a slower pace of model development. Salesforce shares have dropped 8% this year even after a massive pop in August, while Adobe and Autodesk are down significantly more.

On the Dreamforce floor, Salesforce customers and partners told CNBC that older and cheaper AI models are powerful enough for everyday sales and customer service work. Jaya Rohit Vuyyuru, a vice president at consulting firm SummitX, said "The frontier models are way ahead already" while "a lot of the customer base is still getting their feet wet." Tim Sanders, chief innovation officer at software reviewing company G2, said "The majority of agentic outcomes aren't driven by frontier capabilities. They're driven by last year's AI." Kevin Lee, technology chief at cloud contact center software vendor Nice, said his company does not count on high-end models such as Fable for most workloads, adding that even one generation behind is highly performant and effective. Docusign CEO Allan Thygesen said the electronic signature software developer uses all the big frontier models as well as some open-weight models, with larger frontier models reserved for judgment-intensive work such as complex clause analysis, multi-document reasoning and summarization.

Separately, new Lasso research reported by TechRadar found that AI watermarking could unintentionally change how large language models behave after testing Google DeepMind's SynthID-Text. The research found SynthID-Text can change whether models refuse harmful requests, their susceptibility to prompt injection, which tools an AI agent chooses and more. Lasso concluded that the "watermarking procedure can therefore affect both what the model says and what an agent does," referring to the side effect as "sampling drift." Even without an attack, watermarking changed some models' refusal decisions, making them more willing to answer potentially harmful prompts, and prompt injection amplified the consequences. Anthropic recently announced that future generations of Claude would use AI watermarking similar to Google DeepMind's, stressing that a key driver was adherence to the EU AI Act. Lasso urged developers to rerun benchmarks, safety evaluations and other tests rather than apply watermarking blindly to existing configurations, while stressing the findings should not be taken as an argument against watermarking for provenance.