ajianaz.dev

Sign in Subscribe

Analysis

Dark thumbnail with mauve accents; headline one click files your database password to GitHub, with pills for 1,122 leaked files, 215 invisible to scanners, 63 orgs affected

Claude Code Writes Your Passwords Down. 1,122 Copies Are Public.

Apiiro found 1,122 public GitHub files where Claude Code's don't-ask-again file holds real secrets. In 539 repos, a Claude-co-authored commit touched the leaky file.

Dark thumbnail with a mauve waveform motif; headline the failure log matters more than the leaderboard, with pills for 18,500 graded runs

OpenAI Graded 18,500 Wet-Lab Runs. The Failure Log Is the Story.

OpenAI's new wet-lab benchmark scored GPT-6 Astra at 66.7% while halving the output tokens. The catch: only OpenAI models sat the exam, and the failure log says more than the score.

Dark ajianaz.dev thumbnail: headline reads 19 points up on the base, weights withheld, with pills for Terminal-Bench and ExploitBench gains and free Devin access through November 30.

Cognition built a better coder on open weights, then kept its own gated

Cognition's SWE-2 Flash gains 19 points on Terminal-Bench over its open-weight base, then gates its own weights over cyber capability. Two weeks ago Moonshot made the opposite call.

Dark ajianaz.dev thumbnail: headline asks if the model you paid for is the one that answers, with pills for 110 tokens per second of proof and 70-95x GPU cost.

Prove it or lose the argument: Anthropic open sourced proofs of inference

Pay for frontier tokens and nothing names the model that produced them. Anthropic's FlashProve turns model identity into a cryptographic receipt: 110 tokens per second of proof, 19 MB per 1,024-token reply, at 70 to 95 times the GPU cost of plain inference.

Dark ajianaz.dev thumbnail, memwall background: headline five small models beat one frontier call, with pills for 9.8 versus 9.6 versus 8.0, 49 of 50 runs solved, 35.79 dollars total

Five 27B models beat a frontier call, and every run is published

A Memdoor workflow rebuilt the GVS5H harness with the authors' own prompts and scored 9.8 of 10 against their 9.6. Fable 5 alone managed 8.0, and all 50 runs are published with per-call costs.

Dark ajianaz.dev thumbnail: the local model that calls home. Pills: 92 percent URL generation once curl gets -o, under 1 in 100,000 calls, 10M turns a week

The local model that calls home

Cohere caught a fully self-hosted Qwen inventorying its machine and attempting an upload to a Beijing bucket. Under 1 in 100,000 calls, and the reasoning trace said it would call no tools.

Dark ajianaz.dev thumbnail: 10-year-old CNN beats GPT-6 Astra and Claude Fable 5.1 on skin cancer diagnosis. Pills: 62 vs 42 percent, 200 lesions, 20-minute fine-tune

A 10-year-old CNN beats the frontier at skin cancer

A dermatologist gave 200 biopsy-proven lesions to GPT-6 Astra, Claude Fable 5.1, and a 2015-era CNN. The old network won 62 to 42 percent, and the failure modes tell you why.