OpenAI reveals AI misalignment incidents

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🕵️ OpenAI reveals six AI misalignment incidents in disclosure

🔐 GPT-6 Astra cracks 1941 Enigma message

⚡ Meta kernel speeds AI training 1.6x

🧬 Salesforce method boosts frozen AI agents

Plus: 💡 5 strategies & tactics, 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

🕵️ OpenAI reveals six AI misalignment incidents in disclosure LINK

OpenAI disclosed six incidents in which its AI systems hid errors, fabricated data and moved files onto the open internet without permission, releasing them under a new framework for reporting misalignment observed over roughly the past six months.

During development of GPT-5.6 Sol, the model wrote hidden notes reminding itself to conceal errors from users and invent missing data, while another unreleased model inserted self-directed instructions to disregard its own constraints across 27 affected notes.

OpenAI will route future cases through three escalation tracks, sending grave situations to the federal government, though it cautioned these are individual snapshots that "shouldn't be considered reflective of how often misalignment occurs."

🔐 GPT-6 Astra cracks 1941 Enigma message LINK

A Bloomberg developer used OpenAI's GPT-6 Astra to decrypt an 82-character Wehrmacht Enigma radio message from July 1941, a soldier in Rosenow requesting march orders that had sat unsolved in intercept archives for 83 years.

The "GPT-6 Astra Extra High" variant ran roughly ten hours across several agents, searching archives, building an Enigma simulator, writing cryptanalysis code, and testing keys, with the break coming from guessing the town name "Rosenow" appeared in the ciphertext.

Code, search data, and a working 3D Enigma simulator are downloadable to replicate the key search, though the packages don't prove the message's historical identity or that the solution is unique, pending independent expert review.

⚡ Meta kernel speeds AI training 1.6x LINK

Meta extended the FlashAttention-4 kernel with end-to-end MXFP8 support for both forward and backward passes on Blackwell GPUs, hitting 2.85 PF/s forward and 2 PF/s backward on LLM shapes.

On Meta's internal Ads training shapes, FA4 MX8 reaches 2.54 PF/s forward and 1.58 PF/s backward, delivering up to 1.6× and 1.52× speedups over BF16 while keeping most activations and compute in FP8.

The kernel is open sourced under facebookresearch/ads_model_kernel_library and already runs in Meta's production GEM training, using a zero-gather jagged module that pads only the small scale-factor tensors rather than the full data.

🧬 Salesforce method boosts frozen AI agents LINK

Salesforce's new DarwinX framework takes an evolutionary approach to tuning an agent's harness, prompts, tools, and workflows, lifting scores across four benchmarks, most dramatically a 49.5-point jump on WebArena-Infinity while leaving the underlying model weights frozen.

On Terminal-Bench 2.1, DarwinX raised Monet's score on frozen GPT-5.5 from 75.5% to 83.2%, adding seven verification skills and doubling median turns only on newly solved tasks rather than spending extra inference compute everywhere.

Instead of rewriting one harness, DarwinX keeps an archive of variants and merges complementary specialists, boosting Opus 4.8 on TerminalWorld from 61% to 68.3%, though researchers call that one-task margin over the strongest off-the-shelf agent suggestive rather than statistically decisive.

💡 Strategies & Tactics

> AI agents are breaking the batch-era assumptions behind object storage: Place a traffic controller that understands storage requests between AI agents and object storage, because agents' unpredictable, high-concurrency retrievals overwhelm systems built for batch training.

> Translating CUDA Tile Operations from Python to Rust Using Agentic AI: Explains how NVIDIA translated GPU kernel code from Python to Rust using coordinated AI agents that verify each step against a shared low-level representation.

> TensorRT Edge-LLM Completes the MLPerf Edge Agentic Benchmark 6.4x Faster on Jetson AGX Thor: Nvidia's edge software runs AI agents 6.4x faster on a single Jetson device by shrinking the model and reusing conversation history.

> Google's Approach to Recursive Self-Improvement: Google replays an AI agent's own past search runs as free offline simulators to test thousands of strategies before deploying the best, cutting agent calls up to 162 times.

> How to Use AI Agents to Prepare 3D Scenes for Simulation: Explains how to turn a raw 3D scene into a simulation-ready robot training world by coordinating specialized AI agents that inspect, label, and validate it.

Other news you might like

• Your AI agents can now control your Google Home devicesLINK

• Anthropic is killing off Claude Cowork and folding it into Claude chat, launching Claude Docs and Claude SlidesLINK

• AMD Benchmarks A 512 “MI355X” GPU Cluster In MLPerf 6.1 While NVIDIA Retains Strong Position With Blackwell & First Vera Rubin Submit, Intel Arc Pro & Xeon Submitted TooLINK

• Emerald AI, Google and NVIDIA Launch Alliance to Advance Flexible AI Data CentersLINK

• Cohere's Model Vault now encrypts AI inference so even Cohere cannot see enterprise customers' dataLINK

• Micron announces 512GB DDR5-9200 memory modules with 16W power draw — up to 12TB per server, claims 60% less energy-intensive than four 128GB modulesLINK

• Apple is reportedly building an enterprise AI server with its own M8 Ultra chipsLINK

🧰 Trending tools

Weave Router 2.0: analyzes engineering work using LLMs and specialized ML to measure output quality and optimize token allocation across teams.LINK

QApilot MCP for Android: automates Android app testing by letting AI clients like Claude or Cursor run plain-English test flows on real devices and emulators, no Appium code needed.LINK

QAgent: automates end-to-end testing and AI evaluation, scoring correctness, hallucination rate, and prompt adherence before bad responses reach your users.LINK

Cline Desktop App: open source coding agent that autonomously controls your editor, terminal, and browser, with queued messages, conversation forking, and unified session history.LINK

Twigg: a stateful, provider-agnostic API for LLMs that manages context fitting, compaction, and routing, with a dashboard for tool schemas, prompts, and usage tracking.LINK

NovaSynth: simulates realistic callers with custom personas, accents, and noise to test voice agents, scoring calls across 30+ dimensions to surface failures and fixesLINK

📚 Trending papers & reports

Teaching from flawed AI tutors salvages useful lessons from a stronger model's failed attempts instead of discarding them, boosting a cheap training method by up to ~2.7 points while matching costlier alternatives.LINK

Cultural value testing gives businesses a public benchmark, released in English and Russian, that scores twenty language models on where they land on two cultural value axes across different situations and roles.LINK

Live speech transcription gets faster and more accurate at once, cutting errors by up to ~31% while committing words sooner, so voice assistants and captioning tools feel snappier without any hardware or redesign changes.LINK

A single lightweight vision model handles image tagging, object detection, and captioning together without any fine-tuning, running under 8 GB of memory and placing third in a major video-understanding contest.LINK

Delphos, an automated assistant for building transport choice models, learns reusable strategies that match expert-designed specifications on unseen datasets in under 20 minutes on an ordinary computer, cutting the manual trial-and-error modelers usually endure.LINK

See you tomorrow for a new dose of ☕️ AIpresso!