Latest

Jul
31
OpenAI GPT-5.6 price comparison: Luna falls 80% from $1 to $0.20 per million input tokens, Terra falls 20% from $2.50 to $2, and Sol is unchanged.

OpenAI made GPT-5.6 Luna five times cheaper. That does not make every model cheaper

OpenAI cut GPT-5.6 Luna API prices by 80% and Terra by 20%, lowering costs for agent workloads while leaving Sol and subscription prices unchanged.
3 min read
Jul
29
Anthropic open-weight AI policy banner showing a capability threshold applied to open and closed models, alongside its alleged 28.8 million-exchange Alibaba/Qwen distillation campaign.

Anthropic rejects an open-weight AI ban but still wants a capability threshold

Dario Amodei rejects an open-weight AI ban, but wants capability-based safety testing, tighter chip controls, and cites a new Alibaba distillation claim.
3 min read
Jul
27
GrapheneOS phone showing an alleged duress-credential wipe during a CBP inspection, alongside the headline about its first legal test.

GrapheneOS's duress password is getting its first legal test

A US prosecution over a GrapheneOS duress password turns a phone-security feature into a test of data destruction, border searches, and device privacy.
3 min read
Jul
24
Banner showing 179 startups opposing Chinese open-weight AI restrictions alongside the disputed Kimi K3 distillation claim.

179 startups warn a Chinese AI ban would help incumbents, not America

179 startups are asking Washington to preserve access to Chinese open-weight AI models, even as administration's distillation case against Moonshot faces pushback.
4 min read
Jul
22
News banner showing OpenAI models escaping a restricted evaluation sandbox and reaching Hugging Face production, with the OpenAI and Hugging Face logos.

OpenAI's Hugging Face breach shows why AI evaluations need real containment

OpenAI says cyber-capable models escaped an evaluation environment and compromised Hugging Face while looking for benchmark answers. The problem was not the prompt alone, but where the sandbox could reach.
3 min read
Jul
21
Graphic showing GPT-5.6 tracing a WordPress REST API flaw to a pre-authentication RCE for an estimated $25, followed by human verification; versions 7.0.2 and 6.9.5 are marked patched.

GPT-5.6 found a WordPress RCE for $25. The expensive part came next

GPT-5.6 Sol Ultra found a pre-authentication WordPress RCE in a multi-agent research run. The AI part was cheap, verifying and disclosing it wasn't.
3 min read
Jul
18
A working Next.js coding test shows zero antivirus detections while hidden fragments from SVG flag files are reassembled and passed to eval().

Elastic found malware hidden in SVG flags. The working coding test was the real trap

A DPRK-aligned campaign used functional Next.js coding tests to deliver credential stealers and a remote-access trojan from payload fragments hidden in SVG files.
3 min read
Jul
16
Cursor logo beside a warning that a repository-root git.exe can run without a prompt on Windows, plus a seven-month disclosure timeline.

Cursor can run a repository's `git.exe`. It says the repo is the trust boundary

Mindgard reported Cursor's git.exe code execution flaw in December 2025. Seven months later, Cursor calls it out of scope and points to Workspace Trust.
4 min read
Jul
15
PrismML Bonsai 27B benchmark summary showing 3.9 GB binary and 7.2 GB ternary builds, with 11 tokens per second on an iPhone.

Bonsai 27B fits on a phone. But does it run well?

PrismML compressed a 27B-class model into a phone-sized package, but kernels, memory overhead, and real device performance still determine whether it is practical.
3 min read
Jul
13
News banner: npm 12 blocked --ignore-scripts bypasses, but Jscrambler still got compromised in three days. Safe version: 8.22.0.

npm 12 blocked --ignore-scripts bypasses. Jscrambler's compromise found another way in three days

A compromised Jscrambler npm release ran an infostealer from a preinstall hook, then moved the same payload into normal package code once npm 12 disabled scripts by default.
3 min read