Meta AI hacked, stolen Instagram passwords and hijacked accounts

Attackers have seized control of high-profile Instagram accounts by exploiting Meta AI’s support assistant. By requesting to link a new email address to the targeted account, they were able to obtain a password reset link without ever accessing the victim’s email. Meta has since patched the vulnerability, which may have been exploited for several months.

With just a polite request, it became possible to take over a coveted Instagram account. An attacker would initiate a conversation with Meta’s support assistant and essentially state that they needed to associate a new email address with an account for which they provided the ID. The assistant complied, sending them the password reset code. No malware was involved, no traditional software vulnerabilities were exploited, and no access to the victim’s inbox was required; a simple phrase typed into a chat window unlocked the door that, it turns out, was left wide open.

This tactic targeted rare and notable accounts, those that fetch a high price on illicit markets. Perhaps even more alarming than the choice of targets was the duration of exposure. The vulnerability may have been exploited for several months, possibly dating back to February 2026, compromising thousands of accounts before it was revealed in early June. The sole layer of protection that withstood the attack was multi-factor authentication (MFA), which most of the targeted accounts did not have enabled.

A Natural Language Attack Chain

This operation required neither malware nor conventional software vulnerabilities. It exploited the very logic of the assistant. Before engaging in conversation, the attacker would activate a VPN or a residential relay configured to the expected geographical area of the target to avoid triggering fraud detection mechanisms as soon as they logged in. Meta’s support system treated this location as an authorization signal; spoofing this information was sufficient to disarm some of Instagram’s automated defenses.

Next came the clear instruction provided to the assistant: something along the lines of “link this new address to my account; here is my ID, I will send you the code.” The assistant did not impose any strong authentication checks before executing this privileged action. In security terms, the pathway taken by the assistant did not require a mandatory authentication barrier before the sensitive operation could occur. This manipulation is an example of instruction injection, where user input is misinterpreted as a legitimate command rather than data that needs verification.

Notable Accounts Among the Victims

The accounts targeted were not chosen randomly. Among them were the Instagram account of the Obama-era White House, inactive since 2017, the account of Sephora, and that of a senior non-commissioned officer in the US Space Force. The resale value of these rare, vintage, or highly recognizable identifiers explains the attackers’ interest, who preferred monetizable accounts over those protected by stronger defenses.

According to investigations that uncovered the breach, the vulnerability may have been active for several months, possibly as early as February 2026, allowing the compromise of thousands of accounts. Meta has not confirmed this volume. The length of exposure remains one of the most concerning aspects of the incident, as it suggests a design flaw in the support process went undetected while the assistant processed user requests.

Meta’s Fix and the Role of Multi-Factor Authentication

Meta has confirmed that the issue has been resolved. “This problem has been addressed, and we are securing the affected accounts,” said Andy Stone, Meta’s vice president of communications, in a public statement. At this stage, the company has not issued a dedicated press release, and this remains the only official comment. Specific details regarding the timing of the patch deployment and the exact number of accounts affected have not yet been disclosed.

One key takeaway from reports is that accounts protected by multi-factor authentication managed to escape takeover. Even the least robust form, a one-time code sent via text message, would have been sufficient to interrupt the chain at the time of reset. Yet, several reported victims already had an active protection on their Meta account, raising caution about the exact efficacy of this safeguard until Meta provides more details about the conditions of the exploitation.

AI-Powered Automated Support: A New Exposure Surface

Meta rolled out this support assistant in March 2026, touting its capability to resolve account issues end-to-end and directly execute an increasing array of requests within Facebook and Instagram. The incident highlights the risks involved when a conversational assistant gains the ability to act on sensitive operations without independent identity verification prior to executing an action. The distinction between an assistant that suggests and an agent that executes becomes a direct security concern.

For chief security officers, the lesson extends beyond Instagram. Any deployment of an agent capable of modifying account settings, adding an address, or initiating a reset must incorporate identity verification within the execution path itself, rather than relying on the conversational layer controlled by the user. Instruction injection is already among the recognized threats against agentic AI systems; the Meta incident provides a live demonstration on a public service used by hundreds of millions of people.

This episode serves as a reminder that a support assistant does not automatically inherit the safeguards of the interfaces it replaces. Identity checks that protected the human recovery path must be recreated and fortified within the automated pathway. For users, activating multi-factor authentication remains the most effective immediate measure. For organizations deploying support agents, the priority must be to secure privileged actions behind verification that the conversation cannot circumvent.

Scroll to Top