GPT-5.6 Sol Uses Twice the Tokens of GPT-5.5: GPT-5.6 Sol xhigh now uses more than twice as many tokens per session as GPT-5.5 xhigh in my Codex workflow. For Codex users, tokens per session means

GPT-5.6 Sol Uses Twice the Tokens of GPT-5.5

GPT-5.6 Sol xhigh now uses more than twice as many tokens per session as GPT-5.5 xhigh in my Codex workflow. For Codex users, tokens per session means the total token count divided by the number of…

vincentschmalbach.com

Codex CLI Has Responsive Tables in the Terminal: LLMs use Markdown tables all the time. A few months ago, Codex CLI displayed them as ordinary wrapped text. In a narrow terminal, the pipes and dashes

Codex CLI Has Responsive Tables in the Terminal

LLMs use Markdown tables all the time. A few months ago, Codex CLI displayed them as ordinary wrapped text. In a narrow terminal, the pipes and dashes were hard to read. Codex CLI now has responsive…

vincentschmalbach.com

Google Lighthouse Adds Agentic Browsing Checks and Cloudflare Adds AI Traffic Controls: Google Lighthouse and Cloudflare are adding different controls for the same emerging class of automated web visitors. Lighthouse added an experimental

Google Lighthouse Adds Agentic Browsing Checks and Cloudflare Adds AI Traffic Controls

Google Lighthouse and Cloudflare are adding different controls for the same emerging class of automated web visitors. Lighthouse added an experimental Agentic Browsing category. Before Lighthouse…

vincentschmalbach.com

Gemini 3.6 Flash and 3.5 Flash-Lite Ignore Temperature, Top-P, and Top-K: Google's Gemini 3.6 Flash and Gemini 3.5 Flash-Lite ignore temperature, top_p, and top_k, according to Google's current migration guide. Existing Gemi

Gemini 3.6 Flash and 3.5 Flash-Lite Ignore Temperature, Top-P, and Top-K

Google’s Gemini 3.6 Flash and Gemini 3.5 Flash-Lite ignore , , and , according to Google’s current migration guide. Existing Gemini requests could send those sampling parameters…

vincentschmalbach.com

I Tried to Make AI Writing Sound Human by Banning AI Words Through logit_bias: I tried to make AI writing sound more human with logit_bias, an API setting that changes how likely a model is to select specific tokens. I built a bl

I Tried to Make AI Writing Sound Human by Banning AI Words Through logit_bias

I tried to make AI writing sound more human with , an API setting that changes how likely a model is to select specific tokens. I built a blacklist from words that appeared unusually often in AI…

vincentschmalbach.com

A Research Agent Can Leak Private Files Through Its Search Queries: A research agent can leak a private document without uploading it. It can read a detail, turn it into a web search, and send that detail to the search

A Research Agent Can Leak Private Files Through Its Search Queries

A research agent can leak a private document without uploading it. It can read a detail, turn it into a web search, and send that detail to the search provider. That can happen during an ordinary…

vincentschmalbach.com

Qualys Says Claude Mythos Preview Helped Find CVE-2026-64600 in Linux XFS: On July 22, 2026, Qualys disclosed CVE-2026-64600, a race condition in XFS, a Linux filesystem. The flaw is in XFS's copy-on-write path, where the fil

Qualys Says Claude Mythos Preview Helped Find CVE-2026-64600 in Linux XFS

On July 22, 2026, Qualys disclosed CVE-2026-64600, a race condition in XFS, a Linux filesystem. The flaw is in XFS’s copy-on-write path, where the filesystem writes shared data into a new block…

vincentschmalbach.com

Google’s Managed Agents Put Your Runtime in Google’s Cloud: Google’s [Managed Agents](ai.google.dev/gemini-api/d...), now in Public Preview, can give a Gemini agent a Google-hosted Linux sandbox.

Google’s Managed Agents Put Your Runtime in Google’s Cloud

Google’s [Managed Agents](https://ai.google.dev/gemini-api/docs/agents), now in Public Preview, can give a Gemini agent a Google-hosted Linux sandbox. Google documents the default Antigravity agent…

vincentschmalbach.com

An OpenAI Model Tried `kill -9 -1` When Its Shell Hung: On [July 20, 2026, OpenAI described](openai.com/index/safety...) an incident in which an unnamed internal model fo

An OpenAI Model Tried `kill -9 -1` When Its Shell Hung

On [July 20, 2026, OpenAI described](https://openai.com/index/safety-alignment-long-horizon-models/) an incident in which an unnamed internal model found many stuck `find` processes and every shell…

vincentschmalbach.com

An OpenAI Model Rebuilt a Blocked Credential at Runtime: On July 20, 2026, OpenAI [described an internal incident](openai.com/index/safety...) involving an unnamed model f

An OpenAI Model Rebuilt a Blocked Credential at Runtime

On July 20, 2026, OpenAI [described an internal incident](https://openai.com/index/safety-alignment-long-horizon-models/) involving an unnamed model for long-running tasks. It saw that other systems…

vincentschmalbach.com

A Research Agent Can Leak Private Files Through Its Search Queries: A research agent can leak a private document without uploading it. It can read a detail, turn it into a web search, and send that detail to the search

A Research Agent Can Leak Private Files Through Its Search Queries

A research agent can leak a private document without uploading it. It can read a detail, turn it into a web search, and send that detail to the search provider. That can happen during an ordinary…

vincentschmalbach.com

Google’s Managed Agents Put Your Runtime in Google’s Cloud: Google’s [Managed Agents](ai.google.dev/gemini-api/d...), now in Public Preview, can give a Gemini agent a Google-hosted Linux sandbox.

Google’s Managed Agents Put Your Runtime in Google’s Cloud

Google’s [Managed Agents](https://ai.google.dev/gemini-api/docs/agents), now in Public Preview, can give a Gemini agent a Google-hosted Linux sandbox. Google documents the default Antigravity agent…

vincentschmalbach.com

An OpenAI Model Tried `kill -9 -1` When Its Shell Hung: On [July 20, 2026, OpenAI described](openai.com/index/safety...) an incident in which an unnamed internal model fo

An OpenAI Model Tried `kill -9 -1` When Its Shell Hung

On [July 20, 2026, OpenAI described](https://openai.com/index/safety-alignment-long-horizon-models/) an incident in which an unnamed internal model found many stuck `find` processes and every shell…

vincentschmalbach.com

An OpenAI Model Rebuilt a Blocked Credential at Runtime: On July 20, 2026, OpenAI [described an internal incident](openai.com/index/safety...) involving an unnamed model f

An OpenAI Model Rebuilt a Blocked Credential at Runtime

On July 20, 2026, OpenAI [described an internal incident](https://openai.com/index/safety-alignment-long-horizon-models/) involving an unnamed model for long-running tasks. It saw that other systems…

vincentschmalbach.com

AI Can’t Do the Last 20%: I've built a lot of software, both my own side projects and client work as a freelancer. AI tools have made me much faster. Before AI, I spent much mo

AI Can’t Do the Last 20%

I’ve built a lot of software, both my own side projects and client work as a freelancer. AI tools have made me much faster. Before AI, I spent much more time writing each feature and working through…

vincentschmalbach.com

Commercial AI APIs Blocked Hugging Face’s Forensic Analysis: Incident response is the work of reconstructing and containing a breach after it happens. Hosted AI models can act as an elastic analysis layer when a

Commercial AI APIs Blocked Hugging Face’s Forensic Analysis

Incident response is the work of reconstructing and containing a breach after it happens. Hosted AI models can act as an elastic analysis layer when an investigation produces more evidence than…

vincentschmalbach.com

One Leading Dash Turned Tailscale SSH Access Into a Root Session: SSH access is often wider than a security policy intends. Tailscale SSH manages authentication and authorization for connections over a private Tailsc

One Leading Dash Turned Tailscale SSH Access Into a Root Session

SSH access is often wider than a security policy intends. Tailscale SSH manages authentication and authorization for connections over a private Tailscale network. Its identity-based rules can…

vincentschmalbach.com

An AI agent that reviews its own database migrations misses its own errors. A query compiles fine in the model's head, but locks a production table on deploy. Run the generated script in an isolated Docker container. Check the returned JSON against a frozen schema to verify the output.

Asking an AI to build order billing often gets you a bloated OrderService class holding thousands of lines of queries because it defaults to copying public database wrappers. If you define domain types and state transitions first, then prompt the model using those rules, you get much tighter code.

System prompts telling models to stick to provided documents often fail. If a refund policy mentions fourteen days, a model might invent a thirty day policy.

Writing long system instructions to prevent AI coding errors rarely succeeds. Prompts cannot enforce strict technical rules.

AI can build you a form engine in an afternoon. It won't tell you five already exist.

Your CLAUDE.md file is a suggestion, not a guardrail.

An 800,000 pound ticketing system collapsed on launch. A database monolith built over a weekend processed 150,000 sales without a hitch.

Job listings are now hiring for the ability to challenge AI-generated code. When generating code is free, the premium skill is spotting the subtle ways machine output breaks.

AI lowers the time to write code, but proving that code works remains the main challenge.

The ugliest codebase in the company has zero outages. The clean ones do.

Ask an AI model to choose a database for a high-write queue, and it defaults to Postgres. It recommends the most popular choice because that is the safe answer in its data.

The UI looks perfect. The login form has smooth transitions. But when you click submit, the API route sends a UUID payload while the database migration expects an integer. Then the background queue drops the failed job because the error handler in that file was never written.

Imagine shipping a WebGPU shader for a simple dashboard in week one because code generation made it easy.