AI's dark side: Malicious use will lower attack costs, new report warns

A new multi-institutional report is drawing a bullseye on a future where artificial intelligence doesn’t just malfunction, but is actively weaponized. Co-authored by OpenAI in partnership with heavy-hitters like the Future of Humanity Institute and the Electronic Frontier Foundation, the paper breaks down exactly how AI flips the script on global security. The core issue isn’t just sci-fi superintelligence; it’s the grinding, practical erosion of defenses. AI lowers the financial and technical bar for pulling off existing attacks, invents entirely new threats we haven’t had to defend against, and turns the already murky task of figuring out who attacked whom into a total guessing game.

The report doesn’t just sound an alarm; it offers a triage plan rooted in blunt realism. Its first demand is for the AI community to shed any naive pretense that code is neutral. “Surveillance tools can be used to catch terrorists or oppress ordinary citizens,” the authors write, highlighting the deep dual-use dilemma where a content filter can just as easily censor journalism as it can block spam. The paper pushes for concrete, if uncomfortable, safeguards: pre-publication risk assessments for sensitive research and the selective sharing of dangerous findings only within trusted circles. It’s a call to grow up and realize that open-sourcing everything without scrutiny is a gamble the world can’t afford.

A significant portion of the solution set is borrowed directly from the cybersecurity industry’s playbook, which has been fighting asymmetrical digital wars for decades. The authors advocate for institutionalizing ‘red teaming,’ where dedicated experts try to break AI systems before the bad guys do. They also push for investing in technological forecasting to spot threats while they’re still on the drawing board and creating formal channels for researchers to confidentially report vulnerabilities they discover in AI models without fear of legal blowback.

The paper grounds its abstractions in a series of chilling, near-future scenarios. One imagines hyper-personalized AI propaganda targeting a specific security administrator. Another sees neural networks being used to automate the creation of rapidly mutating computer viruses. In a more physical realm, a hacked cleaning robot becomes a rolling bomb aimed at a VIP, while rogue states deploy omnipresent AI surveillance to execute pre-emptive arrests based on predictive risk scores. The authors are now pushing to broaden this conversation beyond Silicon Valley, actively seeking out input from national security experts, civil society, and ethicists to prevent a future where these scenarios escape from the page.