News

MIT’s “PhotoGuard” Protects Images From Unauthorized AI Edits

The technology invisibly alters select pixels to throw off algorithmic AI models.

Published

3 years ago

July 25, 2023

Nour Nasir

mit's photoguard protects images from unauthorized ai edits

MIT CSAIL

As AI continues to develop rapidly, chatbots are gaining the power to create and manipulate images, with Shutterstock and Adobe currently leading the way. Despite the obvious power of AI algorithms, the technology has a few pitfalls, one of which is the unauthorized manipulation of copyrighted artwork and images.

MIT CSAIL thinks it has the answer to this growing problem in the form of PhotoGuard, a new technique that alters select pixels in an image to disrupt AI’s ability to understand what the image is.

The altered pixels are known as “perturbations” and are invisible to the human eye but easily seen by AI bots as they scan the color and position of every pixel in an image. Any edits AI tries to make to a protected image will also apply to the fake pixels, resulting in an unrealistic or broken final image, thanks to PhotoGuard.

Also Read: PicSo Review: A Popular AI-Based Text-To-Image App

“The encoder attack makes the model think that the input image is some other image,” explained MIT student and lead author of the paper, Hadi Salman. “Whereas the diffusion attack forces the diffusion model to make edits towards some target image”. The technique sounds complex but could potentially stop malicious actors from reverse engineering protected images by adding minor edits to circumvent copyright.

“A collaborative approach involving model developers, social media platforms, and policymakers presents a robust defense against unauthorized image manipulation. Working on this pressing issue is of paramount importance today,” Salman said in a recent press release. “And while I am glad to contribute towards this solution, much work is needed to make this protection practical. Companies that develop these models need to invest in engineering robust immunizations against the possible threats posed by these AI tools”.

News

ESET Bets On Specialized Security AI At 0.5% The Training Cost

ESET says NeoCortex can estimate an attacker’s next move, with a kill switch and human oversight built in to keep it from going rogue.

Published

5 days ago

November 4, 2026

Dana Saadeh

eset bets on specialized security ai at 0.5% the training cost

A sign-in anomaly, a suspicious process launch and an unusual connection may look harmless alone. Together, they can describe an intrusion that is hard for security to teams to piece together.

To solve that problem, ESET announced ESET NeoCortex on November 4, 2026: a specialized foundational AI model for cybersecurity. It powers ESET LiveCortex, available within the ESET PROTECT Platform, which connects fragmented security signals and turns them into actionable incidents.

Filtering is the main selling point, with LiveCortex processing billions of raw telemetry events daily, and just 0.00111% of them contributing to incidents presented to customers. The aim is less alert fatigue and more analyst time spent on threats that matter.

NeoCortex is built on world-model architecture, learning how digital environments behave and how events relate. ESET says that lets it reconstruct missing incident context, recognize similar past cases and estimate an attacker’s likely next step – what the company calls “minus-one-day defense,” a chance to intervene before the next stage unfolds. ESET points to Black October, the week leading up to October 25, 2026, when critical security issues affecting PHP, Linux, GitLab and other key systems demanded rapid, prioritized responses, as the kind of moment where that context matters.

ESET’s internal research also suggests a specialized model like NeoCortex can be trained at up to 200 times lower cost than a large general reasoning model, while delivering more accurate, defense-focused results. The company credits optimized training, a focus on cybersecurity vocabulary, lower computing requirements and privately owned hardware.

Also Read: Dubai Says Its Deepfake Detector Needs Just 7 Seconds Of Video

“AI is accelerating both cyberattacks and cyberdefenses. With ESET NeoCortex, we are applying our AI investment to develop specialized foundational cybersecurity intelligence that remains under our control and strengthens, rather than replaces, human expertise,” said Juraj Malcho, Chief Technology Officer at ESET.

As part of the design, NeoCortex forms part of a layered architecture combining specialized models, deterministic controls, correlation logic, validation, auditability and expert human oversight. Safeguards include restricted sandboxes, read-only tools, continuous monitoring, defined usage and token limits, a kill switch and human oversight of critical response decisions. The model is also developed and self-hosted on ESET’s own infrastructure, which the company says reduces exposure to shifts among third-party AI providers, including commercial and geopolitical ones.