The newest Instagram “exploit” is the goofiest I've seen

Exploit and immediate impact

• Attackers could convince Instagram’s AI support agent to send account‑recovery codes to an arbitrary email they controlled, apparently after faking being in the right region (e.g., via VPN).
• Once the email was changed, they could reset the password, revoke existing sessions, and lock out the original owner.
• High‑value and short usernames were targeted; some commenters report being hit, with mixed success in regaining access.
• At least one video demo and multiple independent reports are cited in the thread; some early skepticism gives way to “this really happened.”

2FA and recovery: the real weak point

• Many argue that support and recovery flows have always been the soft spot, even with human agents; social engineering of support has long bypassed 2FA (SIM swaps, registrar hijacks, etc.).
• Some say this exploit bypassed 2FA entirely; others counter that it failed on accounts with any MFA and that separate, paid 2FA‑bypass services were used for the highest‑value accounts. The exact behavior is unclear.
• Broader frustration that 2FA can be removed or overridden by low‑level support (human or AI), undermining its security value.

AI vs “just bad design”

• One camp: this is mainly a catastrophically bad account‑recovery design; AI just automated what a gullible support rep could do.
• Another camp: AI is central — it is more suggestible, scales attacks, has no real judgment, and was given tools built for trained humans without new guardrails.
• Several see this as “vibe‑coded” AI: high‑privilege tools wired to an LLM with minimal threat modeling or end‑to‑end tests.

Critique of Meta’s security culture

• Strong criticism that a trillion‑dollar company shipped a near “zero‑auth” reset flow at all, and that security either wasn’t consulted or was overridden.
• Some note Instagram security historically had sophisticated flows, making this regression especially shocking.
• Others mention Meta’s broader move to AI moderation and support has already caused many wrongful bans and opaque, unfixable decisions.

Suggested alternatives and mitigations

• Ideas floated: stricter flows for high‑value accounts; opt‑in “no recovery without 2FA/passkey” mode; long delays with multi‑channel notifications; physical or notarized identity checks; trusted contacts; better use of government IDs or digital ID for verified accounts.
• Multiple commenters argue that any AI agent must be treated as fully untrusted, with default‑deny tools, tight access scopes, and robust testing akin to public APIs.