Hardening Firefox with Anthropic's Red Team

Claude Opus 4.6 identified 22 vulnerabilities in Firefox, including 14 high-severity bugs, by analyzing the JavaScript engine and nearly 6,000 C++ files. While the LLM demonstrated high efficiency in zero-day discovery and automated patching using "task verifiers," it remains significantly less capable at exploit development, succeeding in only two cases at a high compute cost. Anthropic advocates for integrating LLM-powered patching agents into defensive workflows to accelerate the find-and-fix cycle before model exploitation capabilities improve.

Read more

We might all be AI engineers now

The shift toward AI-assisted development is transforming software engineering from manual coding to high-level architectural design and agent supervision. While LLMs can execute complex logic and accelerate debugging, their effectiveness relies on an engineer's foundational intuition to guide the models and verify output. Ultimately, AI serves as a force multiplier for those who understand system trade-offs, making the ability to decompose problems more critical than the act of writing code itself.

Read more

LLMs work best when the user defines their acceptance criteria first

An LLM-generated Rust rewrite of SQLite demonstrated a 20,000x performance degradation because the model prioritized plausible architecture over critical performance invariants, such as O(log n) primary key lookups. This highlights a broader trend of sycophancy in LLMs, where generated code appears correct and passes basic tests but fails under technical scrutiny due to architectural bloat and inefficient defaults. To mitigate these risks, practitioners must define rigorous acceptance criteria and verify performance metrics rather than relying on the "vibe" of syntactically correct output.

Read more

AI and the Illegal War

Anthropic’s Claude has been integrated into the military’s Maven Smart System to automate target identification and coordinate strikes, reportedly processing 1,000 targets within 24 hours. Despite claims of precision, the use of LLMs in kinetic operations has been linked to significant collateral damage and civilian casualties. Critics argue that these high-stakes military deployments serve as a revenue-propping mechanism for AI firms, even as the technology’s contribution to broader economic growth remains negligible.

Read more

What if AI just makes us work harder?

The provided text is a paywalled Financial Times article titled "What if AI just makes us work harder?" which examines the potential for AI to increase labor intensity rather than reduce it. The page also highlights several high-impact industry developments, including Iranian cyberattacks on Amazon data centers, the introduction of strict new US AI regulatory guidelines, and the cancellation of a flagship data center expansion deal between Oracle and OpenAI.

Read more

Nested Training for Mutual Adaptation in Human-AI Teaming

To address mutual adaptation in human-AI teaming, this work models interactions as an I-POMDP that treats human adaptation as a state variable. It employs a nested training regime where agents learn against adaptive partners from lower levels, preventing the brittle implicit coordination typical of simultaneous multi-agent learning. Results in the Overcooked domain demonstrate improved generalization and performance when paired with novel adaptive partners.

ArXiv

Cybersecurity Data Extraction from Common Crawl

Alpha-Root is a cybersecurity dataset extracted from Common Crawl using single-shot community detection rather than iterative content-scoring. By leveraging 20 trusted seed domains, it identifies high-quality domains directly from the web graph.

ArXiv

Bootstrapping Fuzzers for Compilers of Low-Resource Language Dialects Using LLMs

Germinator is a dialect-agnostic fuzzing framework for extensible compilers like MLIR that leverages LLMs to generate high-quality seed inputs. By combining automatically extracted dialect grammars with pre-trained LLMs, it produces diverse test cases that bootstrap coverage-guided fuzzers without requiring manual seed corpora. Evaluation across 91 MLIR dialects showed a 10-120% increase in line coverage and the discovery of 88 previously unknown bugs.

ArXiv

Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)

Infinity-Chat is a dataset of 26K open-ended queries and 31K human annotations designed to analyze mode collapse and the "Artificial Hivemind" effect in LLMs. The study identifies significant intra-model repetition and inter-model homogeneity across diverse tasks, revealing that current reward models and LLM judges are poorly calibrated to idiosyncratic human preferences. This resource provides a framework for evaluating output diversity and mitigating long-term AI safety risks associated with output homogenization.

ArXiv

Unlocking Python's Cores:Energy Implications of Removing the GIL

Python 3.13/3.14's experimental free-threaded build enables multi-core execution by disabling the GIL, yielding up to 4x speedups and proportional energy savings for parallel workloads. However, sequential tasks face a 13-43% energy increase, and memory usage rises across all categories due to per-object locking and a new allocator. Performance gains are highly workload-dependent, with lock contention on shared objects potentially negating benefits.

ArXiv

Moongate – Ultima Online server emulator in .NET 10 with Lua scripting

Moongate v2 is a high-performance Ultima Online server built on .NET 10, utilizing NativeAOT and source generators to ensure deterministic execution and minimal runtime reflection. It features a modular architecture with a sector-based spatial chunk strategy for efficient world streaming and a Lua-driven scripting subsystem for NPC AI and game logic. The project includes a robust persistence layer using MessagePack-CSharp, integrated HTTP management endpoints, and a comprehensive monitoring stack via Prometheus and Grafana.

Repo

Claude-replay – A video-like player for Claude Code sessions

claude-replay is a community tool designed to transform Claude Code session JSONL logs into interactive, shareable, self-contained HTML replays. It visualizes full conversation transcripts, including user messages, assistant responses, tool calls, and thinking blocks, making AI-assisted development sessions easier to share than bulky recordings or raw logs. Key features include interactive playback, secret redaction, and embeddability, facilitating documentation, demos, and analysis of LLM reasoning.

Repo

Deterministic browser control for AI agents (~90% on Mind2Web)

Chromium is an open-source browser project focused on security, speed, and stability. Developers must follow specific checkout procedures instead of standard git clone and adhere to a product-centric directory structure. Documentation is centralized in docs/README.md, with bug tracking managed via crbug.com.

Repo

Context-compact – Summarize agent context instead of truncating it

context-compact is a TypeScript library that prevents LLM context window overflow by summarizing historical conversation data rather than using naive truncation. It utilizes sequential chunking and a running summary to maintain continuity, while offering configurable policies to preserve critical identifiers like UUIDs and file paths. The package includes utilities for token estimation, safety margins, and automated cleaning of tool-result payloads to optimize summarization performance.

Repo

mcp-recorder – VCR.py for MCP servers. Record, replay, verify

mcp-recorder is a VCR-style testing utility for MCP servers that enables recording, replaying, and verifying protocol interactions. It captures full JSON-RPC exchanges into cassette files, supporting both HTTP and stdio transports for deterministic regression testing and client mocking. The tool includes a CLI for scenario-based automation, a Python API, and native pytest integration to ensure server stability and schema consistency.

Repo