More than 50 framework families now scan Cloudflare customers from thousands of IPs and chain stolen credentials to move laterally. Rate limits and detection built around human-paced attackers now face automated attacks at machine volume.
Adapting to AI penetration testing orchestrators: The Modern Security Model →What matters today
- 01
Read
llama.cpp and ggml reach 1.0 and tie API breaks to major versions
Apps that embed llama.cpp have tracked a moving API. From 1.0, any API or ABI incompatibility bumps the major version, so teams can pin 1.x against a stated contract. The cut itself removes 35 deprecated header functions, so old integrations break once.
llama.cpp v1.0.0 : A stronger foundation for the local AI ecosystem → - 02
Read
Anthropic: users approved 93% of Claude permission prompts
When nearly every prompt is approved, the prompt filters almost nothing. Anthropic instead caps what an agent can reach with sandboxes, VMs and egress controls, and its auto mode hands approvals to classifiers plus an input-side injection probe.
Anthropic research → - 03
Try
browser-use 1.0.4 halves cost per task by batching screenshot pruning
Dropping one old screenshot per step rewrites earlier messages and voids the prompt cache. Pruning in one batch kept the cache and cut BU Bench cost from $0.434 to $0.208 per task at the same score, a change any screenshot-heavy agent can copy.
gregpr07 → - 04
Read
LLM judges pass 10 of 11 tau2-bench runs with forbidden writes
Across 2,160 airline and retail episodes, judges reading transcripts agreed with gold-state scoring only 56.2% of the time. A transcript can read as compliant while the database shows a banned write, so a simple rule-based check on writes catches what the judge missed.
Sungbin Im → - 05
Watch
Mozilla reports 8x more Firefox bugs found and fixed with GPT-6 Cyber
Mozilla uses the model to rank investigations, verify reports, hunt related bug patterns and help write fixes. The 8x year-over-year gain covers fixes as well as discovery, so the extra findings reached patches instead of piling up as unchecked reports.
OpenAI → - 06
Read
Tencent Hy4 leads CWE-Bench-AA at 76% for $0.92 per task
Hy4 also ranks third on the Artificial Analysis Cyber Index at 58%, slightly above Grok 4.7 at xhigh for about one eleventh of its cost per task. Security agents run many tasks across a codebase, so per-task cost sets how much code they can afford to cover.
Artificial Analysis →
