Technology

Most AI jailbreak discussions focus on prompts, but this research looks at images instead.

The Florida International University team developed a technique called JaiLIP that makes tiny pixel-level changes to images that humans largely can’t detect. In testing, the altered images increased harmful outputs from a vision-language model and outperformed previous image-based jailbreak methods.

What caught my attention is that the attack doesn’t rely on prompt engineering. It suggests that images themselves can become an attack vector for multimodal AI systems.

For bridges, the nearest-term opportunity is likely magnetic inspection. My team and I co-authored a review, which has not yet been peer-reviewed, on quantum magnetometers for infrastructure inspection. These sensors identify signals from induction responses, magnetic flux leakage, stress, corrosion and operational currents.

In plain terms, these sensors may help map weak magnetic fields near steel, cables or electrical conductors. Changes or disruptions in these local magnetic fields can reveal hidden rust, snapped wire strands inside a thick suspension cable, or abnormal stress points in the steel before a crack even forms.

https://arxiv.org/abs/2604.03288

Israel's government asked Meta to censor social media content about its ongoing war against Iran, according to internal documents viewed by The Intercept.

Company records show that Israel petitioned Meta to take down Facebook and Instagram posts expressing support for Iran, opposition to Israel, and even depictions of Iranian missile impacts.

The government flagged a variety of materials related to the war, including posts mourning the death of Ayatollah Khamenei following his assassination by the U.S. and Israel on the opening day of the conflict, content supportive of Iran’s retaliatory attacks, and Iranian accounts that shared military analysis and propaganda sympathetic to the Iranian regime perspective.

In some cases, Meta complied with the censorship requests, the records show, though it is unclear on what grounds. Meta maintains that it only removes content as required by law or materials that violate its speech policies.

With a ~40 billion usd hole (netloss) Openai keeps it´s word by staying a nonprofit company 🤣