<p>In Meta, an AI agent went out of control and triggered a serious security incident.</p>
<p>What happened: a Meta engineer used an internal AI agent (similar to OpenClaw) to analyze a technical question on an internal forum. The agent independently, without approval, published a response with recommendations. Another employee followed these recommendations β and this set off a chain of events that led to confidential company and user data being accessible to employees without the appropriate access rights.</p>
<p>For almost two hours, the data was exposed. The incident was classified as Sev 1 β the second most serious level at Meta.</p>
<p>Similar problems are occurring everywhere. AWS experienced a 13-hour service outage after changes made by an AI agent. In China, state-owned companies have banned employees from installing OpenClaw on work devices.</p>
<p>Moreover, in February, the AI security director at Meta described a personal experience: she asked the OpenClaw agent to sort her email with the instruction βconfirm before acting.β The agent ignored this and started deleting emails on its own. Stopping it from her phone was impossible β she had to run to her computer βas if defusing a bomb.β</p>
<p>The only consolation in the Meta incident: at least the agent didnβt pretend to be human β its message was marked as AI-generated.</p>
<p>Autonomous AI agents are a powerful tool, but without reliable constraints, they become a source of risk. And the more access we give them, the higher the stakes.</p>
<p><a href="https://t.me/+OvImEUmA7W5mYTRi">βββββββββ ΠΡΡΠ»ΠΈ Π Π²Π°ΡΠ΅Π²Π° βββββββββ</a></p>
