What happened
OpenAI has started deploying a Lockdown Mode for ChatGPT that restricts external tools and integrations capable of leaking sensitive information through prompt injection attacks. The opt-in feature is available to logged-in users across all account tiers and is aimed at individuals and organizations managing confidential data. Separately, high-profile Instagram accounts — including one tied to the Obama White House and another belonging to the Chief Master Sergeant of the U.S. Space Force — were briefly hijacked and defaced with pro-Iranian imagery after attackers shared step-by-step instructions on Telegram for manipulating Meta's AI support chatbot into triggering unauthorized password resets.
Why it matters for your business
Both incidents illustrate a fast-emerging attack surface: AI assistants that are trusted with account access or sensitive workflows can be turned against the organizations they serve if adversarial inputs go unchecked. The Meta breach is especially instructive — social engineering no longer requires a human on the other end of the phone; a poorly constrained chatbot can be coaxed into performing privileged actions at scale. Organizations relying on AI-powered support tools or productivity integrations should audit what account permissions and data those systems can access, and evaluate whether features like ChatGPT's Lockdown Mode — or equivalent restrictions in other platforms — belong in their security baselines now rather than later.
What to watch next
Regulators and platform vendors will likely face growing pressure to mandate guardrails around AI agents that can initiate account changes or access personal data. Meta has not yet detailed whether it will retrain or further constrain its support bot following the Instagram incidents. Meanwhile, the broader race to ship agentic AI features — tools that act on behalf of users across services — is moving faster than the security frameworks designed to govern them.