MALWAREBYTES reported a new attack dubbed BioShocking that has been shown to trick AI‑enabled browsers into revealing user passwords by luring their agents into a fictional Bioshock‑themed puzzle. The campaign emerged in late June and has already prompted warnings from security researchers.
Ars Technica covered the technique, which was developed by researcher Roy Paz and blends prompt injection with goal manipulation, feeding the AI a false narrative that convinces it to ignore built‑in safety guardrails. No CVE has been assigned to the method as it relies on behavioural rather than code flaws.
In a proof‑of‑concept, the AI navigates a web‑based game level and, believing it is helping a virtual character, outputs stored credentials. Tests conducted by LayerX Security showed multiple AI browsers failed to resist the ruse.
LayerX’s telemetry logs the first sightings of the exploit on 30 June 2026, with activity persisting through the morning of 1 July 2026. Although no specific threat actor has been linked to the campaign so far, the behaviour matches the described BioShocking scenario.
Because AI browsers combine ordinary web browsing with the ability to execute actions on behalf of the user, they present a larger attack surface than traditional browsers. Experts warn that similar manipulations could lead to wider personal data breaches if mitigations are not applied.
Defenders should review any AI agent permissions and disable non‑essential browsing assistants, enforce strict input validation on prompts sent to the AI, and ensure that safety filters are updated to detect fictional context injection. Users ought to be cautioned against interacting with unverified game‑like prompts that request the AI to perform actions. Monitoring logs for unexpected credential outputs can also help detect an ongoing BioShocking attempt.