Microsoft Introduces Multi-Model Project Perception for Cyber Defense

Microsoft introduced Project Perception, a multi-model security system that coordinates red-team, blue-team, and remediation agents across its security products. The company says a configuration using its MAI-Cyber-1-Flash model scored 96 percent on CyberGym and cost almost 50 percent less than its current MDASH configuration.

Those performance and cost comparisons are Microsoft-reported benchmark results, rather than independent measurements. A public preview is scheduled to begin August 3.

Featured source: Ars Technica , Microsoft .

Other Stories

AI Forensics Finds Weak Safeguards in Popular Image-Editing Models

AI Forensics found that seven of the nine top-ranked image-editing models it tested on Hugging Face complied with a simple request to undress an AI-generated person. In a separate seven-day honeypot test, 73 percent of more than 1,000 requests were sexual and 6.7 percent of those targeted a minor; the measurements come from the research group’s tests, while Hugging Face policy prohibits nonconsensual sexual imagery and sexual content involving minors.

Filed from: The Verge , AI Forensics , Hugging Face .

AMD Details 256-Core EPYC 9006 Server Processors

AMD detailed its sixth-generation EPYC 9006 “Venice” server CPUs, built on Zen 6 and TSMC’s 2-nanometer process, with up to 256 cores and 512 threads per socket, 16 DDR5 memory channels, and PCIe 6 connectivity. AMD says production is ramping, while TechRadar reports that the broader platform rollout extends into 2027.

Filed from: TechRadar , AMD .

UK Funding Decision Leaves e-MERLIN Support Ending in 2028

UK Research and Innovation’s prioritization removes future support for e-MERLIN, the seven-telescope radio array operated from Jodrell Bank, after current funding ends in 2028. Jodrell Bank and the University of Manchester say they will seek other support for e-MERLIN and the Lovell Telescope; the observatory itself is not scheduled to close in 2028.

Filed from: The Register , UKRI , Royal Astronomical Society .

Court Narrows Google’s DMCA Case Against Search Scraper SerpApi

A federal judge dismissed Google’s DMCA claims where its anti-bot system controlled access to search results that contained no copyrighted material. Claims involving copyrighted components were dismissed without prejudice because Google had not alleged that copyright owners authorized the access control; Google has 21 days to amend that portion and says it plans to do so.

Filed from: Ars Technica , court order .

Biotech Groups Begin Developing Measles-Specific Treatments

Several companies and academic groups are developing monoclonal antibodies and antivirals for measles as US cases rise, but none of the candidates described has entered human testing. Invivyd is preparing data for a clinical-trial application, while antibody and antiviral programs at La Jolla, Vanderbilt, and Georgia State remain preclinical; vaccines remain the established preventive measure.

Filed from: Ars Technica .

Perplexity Brings Its Personal Computer Agent to Windows

Perplexity released Personal Computer for Windows to Max and Enterprise Max subscribers, allowing its agent to work across local files, Microsoft 365 applications, and the web. The company says the product asks before sensitive actions such as sending email or deleting files and does not train on company data; those privacy and control claims are vendor-reported.

Filed from: The Verge , Perplexity .

Fish Audio Raises $50 Million for Voice-Model Development

Fish Audio raised a $50 million seed round led by Coreline Ventures and Capital Today after reporting 8 million users and $21 million in annual recurring revenue. The startup has automated takedowns for voices uploaded without consent, but unapproved copies can remain available until a rights holder discovers and reports them; the usage and revenue figures are company-reported.

Filed from: TechCrunch , Fish Speech .

Starship’s Latest Test Leaves Rapid Heat-Shield Reuse Unresolved

SpaceX’s 13th Starship test flight completed a controlled reentry, but post-flight imagery showed streaking at tile boundaries and signs of tile damage. Three space-industry specialists told Ars Technica that the ceramic-tile design may still require substantial inspection and refurbishment, leaving rapid upper-stage reuse unproven.

Filed from: Ars Technica .

Flag Game Study Finds Multi-Agent Performance Can Decline at Scale

Harvard and NTT researchers found non-monotonic scaling in a controlled “Flag Game” where agents combine partial visual evidence to identify a country flag. Performance improved as the population grew to 16 agents, then polarization and competing camps reduced consensus quality; the result comes from a synthetic workshop benchmark, not an enterprise deployment.

Filed from: The Register , OpenReview .

Thea Energy Receives $20 Million for Fusion-Magnet Manufacturing

Thea Energy received a $20 million ARPA-E SCALEUP award to establish US production lines for modular high-temperature-superconducting magnets. The magnets are intended for the company’s planar-coil stellarator design; the award supports manufacturing scale-up rather than a commercial fusion deployment.

Filed from: TechCrunch , Thea Energy .

From the Community

Apple Patches More Than 60 Flaws in macOS Tahoe 26.6

Apple’s macOS Tahoe 26.6 security notes list fixes across more than 60 entries, including issues involving sandbox escapes, privilege escalation, kernel memory corruption, and privacy controls. Many of the reports are credited to external researchers.

Filed from: Apple Support .

SlopCodeBench Exposes Defect Accumulation in Long-Horizon Coding

A SlopCodeBench evaluation tested models on multi-checkpoint programming tasks where requirements are revealed incrementally. Anthropic’s Opus 5 reached a 24 percent strict pass rate on the tested subset, compared with 17 percent for Opus 4.6, while every evaluated model accumulated defects and none completed a challenge defect-free; the report is a subset evaluation rather than a full benchmark run.

Filed from: HumanLayer .

FeyNoBg Expands an Open Background-Removal Model

Feyn released FeyNoBg, a 263-million-parameter background-removal model based on BiRefNet, together with the NoBg training and inference library. The project reports leading results on four of eight evaluated benchmarks and performance within two percent of the leader on the remainder.

Filed from: Feyn .

AllYourCodebase Packages C and C++ Projects for Zig

The AllYourCodebase organization maintains build.zig packages for projects including zlib, FFmpeg, libxml2, and BoringSSL. Its repositories either wrap upstream releases or add Zig build scripts, aiming to make cross-compilation less dependent on project-specific Make and CMake environments.

Filed from: GitHub .

Yap Uses Apple’s On-Device Speech APIs for Dictation

Yap is an open-source macOS dictation application built around Apple’s SpeechAnalyzer and SpeechTranscriber APIs. It requires macOS 26 or later and keeps transcription on the device without an API key or separate model download.

Filed from: GitHub .

Nix Pins a Rust and BPF Systems-Software Toolchain

Javier Honduvilla Coto describes using Nix to pin the Rust, LLVM, and BPF dependencies of a systems project, making toolchain changes reproducible and easier to bisect. The same setup exposed undefined behavior when a newer compiler introduced stricter checks.

Filed from: Hondu .