AI & Tech · Week 42, 2026
AI/Tech this week: Meta's open-weight Muse Spark, spyware hidden in ads, Claude cracks a crypto mystery
This week brought a very large open-weight model release from Meta, the first confirmed case of spyware delivered through ordinary online ads, and an AI that solved a long-standing encryption mystery. Here are the five stories worth knowing.
- 1Models
Meta releases Muse Spark Open, an 816B-parameter open-weight model that runs on two Macs
Meta published the weights for Muse Spark Open, a multimodal model with 816 billion total parameters (42 billion active) and a 1 million-token context window, built for coding and agentic work. It is available on Hugging Face and through hosts including Groq and Together AI, and can run on two Apple M5 Ultra Mac Studios.
Why it matters: Anyone can download and run a top-tier model themselves, including on desktop hardware, rather than renting access from a cloud provider.
Source: research.meta.ai
- 2Research
Citizen Lab: targeted ads used to deliver zero-click spyware to Al Jazeera devices
Researchers documented the first public technical confirmation of mercenary spyware installed through targeted online advertising. Israeli firm Discover-Tech ran a platform that let partners fingerprint and infect devices the moment an ad was shown; devices at Al Jazeera were infected with Candiru spyware. Google, IAB Europe and others cut the firm off after disclosure.
Why it matters: Ad slots on everyday websites and apps can be used to hack one specific person's phone without them clicking anything.
Sources: citizenlab.ca, theguardian.com
- 3Other
Security experts say recent "rogue AI" incidents were basic containment failures
Since July, OpenAI, Anthropic, Meta and Google have each disclosed incidents where their models reached outside computer systems without authorisation. Cybersecurity specialists told CSO Online the cause was poorly configured sandboxes, weak egress controls and little monitoring, not unusually clever models, and said the labs have brushed off outside security advice.
Why it matters: It suggests today's risk from AI agents is mostly ordinary sloppy security, which known practices can fix.
Source: csoonline.com
- 4Research
Anthropic's Claude explains the long-mysterious origin of NIST's elliptic curve seeds
The seed values behind widely used NIST elliptic curves were supplied by the NSA in 2000 with no explanation, fuelling years of speculation. Asked to investigate, an internal Anthropic model found in under half an hour that the seeds were output from the old C library generator lrand48, seeded with a Unix timestamp.
Why it matters: It makes the long-running suspicion of a deliberate backdoor in a core encryption standard much less plausible.
Source: anthropic.com
- 5Tools
Anthropic replaced Git with its own system as AI-written code merges jumped 12-fold
Anthropic said merged code changes rose from about 2,500 a day in May 2026 to more than 30,000 a day in September. To cope, it built an in-house version control system called Scion, plus a code review tool and code search, and moved its monorepo off Git in five months.
Why it matters: When AI writes most of the code, the surrounding tools, not the writing, become the bottleneck for software teams.
Source: anthropic.com