The dark side of AI has been a hot-button issue for polarized camps who argue about the degree of danger posed by AI. The debate has gained new momentum as fresh controversies involving AI safety and rogue AI agents have come to the fore.
Most recently, security concerns caused regulators on multiple fronts to take action: The EU enforced its AI Act Article 50 transparency obligations, China shut down AI companion apps serving 345 million users overnight, Colorado's AI Act became enforceable, and the White House pressed OpenAI to slow the release of a new model over safety concerns.
OpenAI itself is now the subject of a multistate subpoena examining chatbot user safety ahead of its IPO, and courtroom testimony from a former safety-team employee has described the company's culture shifting from research-first to product-first over time.
The rapidly increasing power of AI models has triggered security concerns about the ability of AI to breach defenses and arm attackers with ever more sophisticated defense penetration capabilities. Security concerns caused Anthropic's Claude Fable and Claude Mythos to be suspended for three weeks by the U.S. Department of Commerce export controls before access was restored under restrictive conditions.
AI-Related Injuries and Deaths
Besides national espionage and corporate security breaches, AI has been linked to an assortment of personal injuries. Multiple wrongful-death lawsuits have been filed against OpenAI and Character.AI by families alleging that chatbot interactions played a role in their teenagers' suicides. Google and Character.AI reached settlements in January 2026, and the companies have since restricted or removed chat access for minors in response.
Hallucinated AI outputs have caused financial and reputational damage, a pattern serious enough that courts, regulators, and enterprises are now treating "the model said something confidently false" as a liability category rather than a quirk to shrug off.
Trouble in Agentic Land
Surveys show that 65% of organizations have already had at least one cybersecurity incident caused by an AI agent in the past year. Thirty-five percent of executives admit they aren't confident they could shut a rogue agent down before it did damage.
Those two numbers come from completely different studies, run by different researchers and on different populations. The data shows the same types of failures and concerns coming from two different sources. They also show how widespread the problems are and how unprepared organizations are to deal with the issues.
It Gets Worse
Five distinct bodies of research published in the past four months describe the same underlying pattern from five different angles: governance, tool security, shadow usage, operational damage, and the toll AI is taking on executives forced to manage the deployment of AI within their organizations.
1. Mixed predictions for agentic outcomes
Gartner's agentic AI research shows agentic AI growing rapidly in adoption while suffering significant failure rates in practice. The firm predicts that 40% of enterprise applications will have embedded AI agents by the end of this year, up from less than 5% in 2025, an eightfold jump in 12 months. However, Gartner also predicts that by 2027, 40% of enterprises will demote or decommission those same agents because governance gaps are discovered only after a production incident.
A separate Gartner prediction puts an even higher number on project cancellations: More than 40% will occur by the end of 2027, driven by escalating costs, unclear business value, and inadequate risk controls. Gartner's explanation for the failure rate: Enterprises treat agent governance as binary, "either locked down or fully trusted," and that binary choice is the root cause of failure.
Deloitte's survey findings show an even wider gap between deployment and readiness: Nearly three-quarters of companies plan to deploy agentic AI within two years, but only 21% currently have a governance model mature enough to handle it.
2. Agent impersonators penetrate your browser's defenses
On July 18, a security weakness in Anthropic's Claude for Chrome extension was disclosed: Under the right conditions, another browser extension could trigger a predefined AI workflow without a genuine user click, potentially exposing information in Gmail, Google Docs, and Google Calendar, or initiating actions in connected services like Salesforce.
The threat isn't a break-in. It's impersonating a trusted user, an unauthorized process that simulates the security verification that a genuine user's click would normally establish.
This is not an isolated flaw. In July 2025, malicious code was committed into the Amazon Q Developer extension for VS Code and shipped in a public release before removal. Independent security analysis found the embedded prompt was written to make the AI model produce commands that would wipe local files and cloud resources. In both cases, the AI model wasn't tricked into doing something unusual. It was given a plausible-looking instruction and followed it, which is precisely what an AI assistant with system permissions is designed to do.
3. AI hidden in the shadows
Verizon's 2026 Data Breach Investigations Report showed that shadow AI detections rose fourfold in a single year, with 45% of employees found to be regular AI users on corporate devices, whether authorized or not.
Zylo's 2026 SaaS Management Index found that 77% of IT leaders discovered AI-powered tools operating in their environment without their knowledge.
Mimecast's State of Human Risk 2026 report, which surveyed 2,500 IT security leaders across nine countries, found that 80% of organizations were concerned about data leaking through generative AI tools, yet 60% had no specific strategy to address the problem.
Netwrix survey findings showed that organizations in which AI significantly expanded the number of identities with access to company data had a 43% breach rate over the prior year, versus 11% for organizations where access was not expanded. That's a fourfold difference tied to one variable, how much access was handed out without anyone tracking it.
These findings indicate that shadow AI might be creeping in at higher rates while being more difficult to detect than old-fashioned shadow IT ever was. The entrance of unauthorized software into an environment used to require someone who knew how to build it and deploy it. Shadow AI just requires a browser click that installs an unconscious agent.
4. Agents run wild
Cloud Security Alliance and Token Security's joint research, published in April, found that 65% of organizations experienced at least one cybersecurity incident caused by an AI agent operating on their network in the past year. Of those incidents, 61% involved sensitive data exposure, 43% caused operational disruption, and 41% resulted in unintended actions across business processes.
Their framing is worth repeating: The agent isn't malfunctioning. It's doing exactly what its permissions allow, which means this is a governance failure, not a technology failure.
Separately, Cyera analyzed more than 7,200 publicly reported AI-security and operational incidents and verified 344 enterprise-relevant cases of agent-inflicted damage since September 2023, including 188 where an autonomous AI agent caused direct organizational harm with no external attacker involved.
Three incidents from that report are worth looking at in detail:
5. AI Anxiety at the Top
Writer's 2026 enterprise AI adoption survey adds a human dimension to the technical and incident-related findings of other surveys. The survey found that 29% of employees, 44% of Gen Z specifically, admit to actively sabotaging their company's AI strategy by entering company data into public tools, using unapproved tools, or simply refusing to use AI at all. Seventy-six percent of executives called that a serious threat to their company's future. However, 60% also said they planned to lay off employees who won't adopt AI.
Sixty-seven percent of executives said they believed their company had already suffered a data leak or breach from an employee using an unapproved tool, which coincides with the shadow AI research cited above.
The survey also showed contradictory feelings within the executive suite about the effectiveness of AI. Ninety-seven percent of executives said that AI has been beneficial overall, while 48% said AI adoption by their company had been a massive disappointment.
Thirty-eight percent of CEOs reported high or crippling stress about the effectiveness of their AI strategy, while 64% said they feared they could lose their job if they failed to lead the transition well.
Sources: AI Pulse · Where This Breaks · workplaceai.ai. Introduction sourced from: OpenAI's multistate subpoena, PBS/AP, June 13, 2026; the White House's request to slow a model release, TechCrunch, June 25, 2026; former OpenAI safety-team testimony, TechCrunch, May 7, 2026; the Claude Fable/Mythos export-control suspension and restoration, per Anthropic's public statement; and reporting on AI-chatbot wrongful-death litigation and the Google/Character.AI settlement from CNBC and Gulf News, January 7, 2026. This is a sensitive area of ongoing litigation; details here are limited to the public, procedural facts, filings, settlements, and resulting policy changes, not case specifics. Governance data from Gartner's press releases of August 26, 2025 and May 26, 2026, and Deloitte's agentic AI governance research. Claude for Chrome disclosure: The Cyber Security Hub, July 18, 2026. Amazon Q Developer incident: Independent security analysis as reported by ODSC, June 2026. Shadow AI research: Verizon's 2026 Data Breach Investigations Report, Mimecast's State of Human Risk 2026 report, Zylo's 2026 SaaS Management Index, and Netwrix's 2026 Data and Identity Security Report. Agent-incident research: Cloud Security Alliance and Token Security's "Autonomous but Not Controlled," published April 21, 2026, and Cyera's independent analysis of 7,200+ publicly reported incidents, published June 2026. PocketOS incident as reported by The Guardian; Air Canada incident as reported by CallSphere; AWS incident as reported by Reuters. Workforce data: Writer's 2026 enterprise AI adoption survey. Every figure above is attributed to its original researcher; none is a WorkplaceAI study. Given how fast this research is being updated, verify current figures before citing them elsewhere.