OpenAI turns prompt-injection testing into a system
GPT-Red shows OpenAI is operationalizing automated adversarial testing with self-play, not treating prompt injection as a one-off eval problem.
Technical AI signals · Daily at 8 PM ET
Wednesday, July 15, 2026
Today split between OpenAI productizing automated prompt-injection testing, a fresh wave of open long-context/open-weight model drops, and more evidence that agent harnesses and connectors are now the real attack surface.
GPT-Red shows OpenAI is operationalizing automated adversarial testing with self-play, not treating prompt injection as a one-off eval problem.
The strongest open-model signal today is capability density: 1M-context models, large MoEs, and agent-oriented positioning are converging into practical long-horizon stacks.
Thinking Machines paired an open multimodal release with NVIDIA-backed training details and a downloadable checkpoint, which is what operators need to test, fine-tune, and actually run it.
MCP-connected Claude artifacts and Codex pulling from Drive, Slack, and local files point to the same shift: assistants are moving from chat into cross-app execution layers.
Claude web_fetch exfiltration, destructive-command blocking, and new writing on harness-level control all reinforce that the system around the model is now the security boundary.
This X post only contains a link with no additional text.
Allen AI says SciArena collected about 3,900 votes from about 1,700 users before retiring on July 15.
The post says Odyssey built a multiplayer world model with learned simulation and rendering and demoed GoldenEye deathmatch streamed to multiple players.
This is a repost saying LingBot-VLA 2.0 is open source and trained across diverse robot configurations.
This is a repost describing a new Microsoft paper on debugging agent trajectories at scale.
The post argues that infra and developer-experience automation now also speeds up an 'army of agents,' in addition to speeding up the engineer.
The post says the /goal implementation plus prompts for compaction, orchestrator, and subagents are open source.
The post claims Thinking Machines, Microsoft MAI, Gemma, and gpt-oss all use sliding-window attention.
The quoted line says learned relative positional embeddings performed and extrapolated better to longer sequences than RoPE.
The repost says ripgrep 15.2.0 includes gitignore bug fixes and a speed-up for large directory trees.
The talk is framed around serving a growing user base on the same hardware when GPU procurement takes months or years.
This is a Hugging Face blog post about lessons from building Shippy.
This is a Hugging Face blog post about model routing.
The authors say their monitor-side NLA readouts surfaced reward-hacking knowledge better than direct verbalized judgments on some datasets.
The post is about building multi-camera 3D object tracking across large spaces with DeepStream 9.1 skills.
NVIDIA says CUDA 13.3 adds support around carryless multiplication, a primitive x86 CPUs have shipped for more than 15 years.
This is a Google Research post in Algorithms & Theory on diffusion model creativity.
The post says ablations tie a chess transformer's knight-fork decodability to one specific attention head after earlier evidence around block 5.
The top comments argue xAI opened Grok Build for tactical reasons and say no independent data-destruction certificates have been presented.
Commenters report running Qwen3.6-35B-A3B on a 16GB Mac at 7–9 tokens per second and debate whether provider token costs are lower than local electricity costs.
Commenters suggest agent-facing APIs should include a feedback tool, while another commenter argues against removing defaults because it hurts human debugging with curl.
Commenters say many developers build similar memory systems using markdown storage plus read, write, update, and search endpoints.
A top commenter asks how Coasty differs from other computer-use agent APIs in the same YC batch and prior batches.
Commenters ask about execution-time and storage quotas, pricing, and whether the included SQLite database has size limits.
Commenters say MikroTik suits LLM-assisted configuration because it publishes markdown reference docs, can dump full configs to one text file, and keeps CLI commands stable across versions.
The top comment says the hard part of AI root cause analysis is no longer the model.
The paper proposes function-aware FIM mid-training, using program dependency graphs and complexity-inferability selection to teach models the action-observation-continuation pattern found in coding agents.
This paper introduces ACQUIRE, a QA-driven framework that targets repository knowledge gaps before attempting a fix.
SynthDocBench is a fully synthetic benchmark that independently varies document length, layout, modality composition, and question type to isolate failure modes.
MonkeyOCRv2 pretrains on 113 million document images across 17 languages and combines image-to-text generation with pixel-level document reconstruction.
SpectraReward turns pretrained MLLMs into training-free reward models by scoring how well the original prompt can be recovered from a generated image in a teacher-forced forward pass.
The work introduces SearchGen-20K and SearchGen-Bench with 20,839 prompts across 12 failure categories and 22 domains, where frontier open generators score 21 to 28 out of 100.
Blind-Spots-Bench contains 235 student-sourced questions designed to expose simple-for-humans failures that standard benchmarks may miss.
The authors say representation-pooling probes on intermediate activations gave better calibration than chain-of-thought outputs on Eternis-Forecaster 8B and also held for GLM-4.7-Flash and GLM-4.5-Air.
AMID is a multi-agent framework for medical imaging ML that adds Data-Conditioned Method Planning to turn coarse task search spaces into executable, parallelizable method lanes.
MuScriptor combines synthetic-data pretraining, real-audio fine-tuning, reinforcement learning, and instrument-presence conditioning for multi-instrument transcription.
This podcast focuses on working around LLM context limitations, with Dex Horthy discussing what context engineering does and does not solve today.
OpenAI demos turning ChatGPT workflows into scheduled tasks, including a daily brief, feedback triage, sharing with Workspace Agents, and converting a workflow to a cloud agent.
This episode includes Lovable CEO Anton Osika discussing one million new apps a week and lower builder costs alongside a separate interview with former Intel CEO Pat Gelsinger.
OpenAI demos a customizable keyboard for Codex that handles voice dictation, task switching, reasoning controls, permission requests, and queued follow-up prompts.
OpenAI's demo shows Codex using a red-card trigger to exit a meeting, block a calendar, set an out-of-office reply, and log the user out of a computer.
This episode outlines five trends including harnesses, loops, skills, and software factories, and frames the shift as better human control rather than unchecked autonomy.
This 0.9B audio model does single-pass transcription, diarization, timestamps, and acoustic event awareness for recordings up to 90 minutes across 50+ languages.
Unlimited-OCR targets one-shot long-horizon parsing and the repository says it added vLLM inference support on 2026-06-28.
OvisOCR2 is a 0.8B page-level document parsing model that outputs Markdown in reading order and was post-trained from Qwen3.5-0.8B using SFT, RL, and OPD.
This dataset releases 4,665 converted Pi trace sessions from 60 source sessions, with 3,799 tool actions and 866 assistant text actions.
EdgeBench evaluates agent learning over time with 134 real-world tasks, 51 publicly released tasks, and more than 38,000 hours of agent interaction.
This dataset supports Vera, a layered diffusion framework that jointly generates an edit layer, an alpha matte, and a composite video for content-preserving video editing.
This repo packages small, composable agent skills intended to work with any model rather than owning the whole engineering process.
Open Interpreter says it is a Codex fork focused on harness emulation for low-cost models and includes browser and native-app computer use plus native sandboxing.
Hallmark says it applies one of 20 themes, 4 verbs, 57 slop-test gates, and a pre-emit self-critique to generate less templated UI.
This repo collects 100+ open-source AI agents, agent skills, and RAG apps with templates that work across Claude, Gemini, GPT, DeepSeek, Llama, and Qwen.
The project presents a trading agent with API and MCP support and includes a warning that specific X, Virtuals, and token assets are not official.
Project AIRI aims to recreate Neuro-sama-style virtual characters and provides installation paths including winget, Scoop, and Homebrew Cask.
DeepTutor's recent releases add single-document removal from a knowledge base and multimodal image extraction through LlamaIndex ingestion.
This is an open online textbook covering math, computer science, and AI from the ground up with an intuition-first approach.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.