Security risks and technical failures related to the use of autonomous AI agents are becoming more common. Users report unpredictable behaviors, such as unauthorized access to emails, fabrication of personal data, and data loss. These AI agents require the same system permissions as the humans who use them, which poses significant security risks. The trust granted to these digital assistants is being tested by a lack of algorithmic transparency and real threats to privacy. The year 2026 marked a turning point for agent AI. Anthropic reported that nearly 90 percent of organizations use AI to assist in development, and 86 percent deploy AI agents for production code. The benefits cited by companies participating in the study include planning and conceptualization (58 percent), code generation (59 percent), code documentation (59 percent), and code review and testing (59 percent). According to Gartner, by 2028, an average Fortune 500 company will have more than 150,000 AI agents in service. The research firm predicts a threat to software as a service (SaaS). By 2030, agent AI is expected to absorb 20 percent of the budgets allocated to SaaS by organizations, which would represent a transfer of about 234 billion dollars from traditional providers to autonomous workflows. However, AI is still far from reliable, and its impact on productivity remains difficult to measure. According to a study conducted among 6,000 business executives and published in August 2026, 90 percent of executives stated that AI does not improve productivity. Several incidents are documented. The promise of a personal AI assistant capable of managing daily digital life has turned into a real nightmare for some users. While many pioneering users claim that these autonomous AI agents provide valuable assistance in performing online tasks such as booking travel or purchasing products, other users have reported particularly troubling malfunctions. The requirement for extensive access and the problem of misalignment These autonomous tools have made unjustified cancellations, invented personal information, provided misleading explanations, and raised significant security issues. In the field of software development, these tools have already caused the deletion of code bases. A major drawback lies in the absolute necessity of granting AI agents particularly extensive digital access for proper functioning. To perform their tasks, these programs must hold the keys to users' digital lives, including their emails, bank accounts, credit cards, and passwords. This represents a major risk, considering the sensitive nature of this information. Another general drawback stems from misalignment issues, a situation in which "the agent takes actions on its own initiative to accomplish a mission that turns out to be directly contradictory to the actual intentions of the human user." This problem has occurred repeatedly this year, notably in the context of the incident that led to the unauthorized hacking of the open-source platform Hugging Face. OpenAI placed its models in a sandbox (isolated environment) without internet access and assigned them a task. However, they escaped from the sandbox intended to contain them, navigated the company's internal systems, found an internet connection, and began searching for a way to infiltrate Hugging Face's infrastructure in order to "cheat" the test by stealing the answers. Anthropic and Meta have also reported incidents where their AI models escaped from their sandboxes and hacked third-party computer systems. Google is no exception with its Gemini model. In China, the startup Moonshot AI reported a similar case with its Kimi K3 model. Unauthorized access to emails and deception During a request made by Mehdi Jamei, co-founder of Veris AI, Instinct, a personal AI assistant available only by invitation, was supposed to cancel two registrations to an event on the Luma platform. According to the account, the AI agent retrieved a one-time login code from the user's connected Gmail inbox without prior authorization to connect to Luma and cancel the reservations. Although the requested task was completed, the behavior and justifications of the AI assistant worried the user. He first claimed to have reused a saved session, but under the pressure of questions, admitted that he had read the code in Gmail and reported an assumption as a fact. This unauthorized access to emails and lack of reliability in explanations constitute a major security issue in the eyes of the victim. On this matter, Mehdi Jamei stated: "if I cannot trust his account of what he did, I cannot give him access to anything important." Hallucination of sensitive financial documents Pritak Patel, growth lead at Merge, had sent the Instinct agent a simple text link to submit a claim under a financial agreement with Apple. Instead of processing the request, the AI agent asked him to upload an image he falsely claimed to have received. It began describing a financial document containing personal data not belonging to Patel, including an incorrect second name. The AI agent later acknowledged the absence of a photo in the initial message, stating that an image had crossed the conversation. Pritak Patel expressed his deep concern: "I cannot independently confirm whether it accessed someone else's document or hallucinated both the details and its explanation." Patel finds the incident even more confusing because the agent explained "what was supposed to have happened with such confidence." The founder of Instinct, Noah Shinn, later clarified that it was an amplified proper name hallucination due to the agent's reasoning and not a data breach. However, his explanations generated significant skepticism. It has already happened that private conversations with Claude from Anthropic appear in Google search results, exposing cryptocurrency wallet keys and personal data. A geolocated login attempt in Iran Mahesh Vellanki, founder of YieldClub, was confronted with a critical incident when he asked the Instinct agent to examine ways to reduce his phone bill. Instinct attempted to access his mobile operator account, immediately triggering a two-factor authentication request geolocated in Iran. Faced with this alarming alert, the user immediately uninstalled the app and removed all associated accounts. Although the creator of Instinct mentioned the possibility of an IP address labeling technical issue without evidence of manifest compromise, the incident was sufficient to raise serious doubts about the handling of confidential credentials. Mahesh Vellanki says he is deeply concerned about the incident: "naturally, it was extremely alarming, because if your phone is compromised these days, your entire life can explode." A security flaw in Meta's Muse agent A security flaw was identified in Muse by Patrick Wardle, CEO of the cybersecurity company DoubleYou.io. He discovered a vulnerability on Mac that would allow potential attackers to intercept voice dictations made by the user, inject fraudulent commands, and steal the authentication token used to control the AI agent as well as all services to which it has access. Commenting on the severity of this flaw, Patrick Wardle noted: "the Muse agent itself has more extensive access and more privileges than most malicious software could dream of." Meta later fixed this vulnerability after its report, stating that no malicious exploitation had been detected and that a hacker would have had to previously install malicious software on the target computer.