Security researchers used Claude to exploit a heap buffer overflow in the libheif image library to gain remote code execution on OpenAI's Discourse forum, then leveraged an SSO misconfiguration to take over employee ChatGPT accounts and access internal repositories. The entire attack chain—from vulnerability discovery to proof-of-concept in OpenAI's internal monorepo—took less than 72 hours and cost under $3,000 in API tokens.

• Claude Opus 5 cracked a reliable x86-64 exploit for the libheif vulnerability within hours of its release, while Opus 4.8 had failed across multiple sessions to do the same with ASLR enabled
• The researchers demonstrated they could access any OpenAI service using the company's SSO system, not just Discourse—meaning compromising any connected third-party service would grant similar access
• Across a broader two-month research campaign targeting Slack, Meta, GitHub, and others, AI models adapted the same libheif exploit to different environments in one or two days each, with only Shopify detecting the activity despite thousands of malicious image uploads

This issue rounds up dev tools and research, from a zero-latency domain autocomplete engine and Transformer internals to Go’s padding trick for faster clears. It also covers memcached vs Redis, using AI for large code diffs, building desktop apps with Deno, orchestrating agents with Orca, and GLM-5.2’s performance plus its head-to-head with Claude Opus.

• Wirewiki's autocomplete handles 240M domain names with p99 latency of 0ms by caching popular domains client-side and only querying the backend for unfamiliar entries.
• GLM-5.2 beats most open models on benchmarks but lacks vision checks and can overfit, making it better for cheap text-heavy work.
• Claude Opus 4.8 is slower and pricier than GLM-5.2 but produces cleaner code and visual feedback, winning a 3D WebGL platformer test.

The article argues that AI “loops”—self-prompting agents using a goal, context, evaluation, and an agent—outperform single-shot prompts for long-running tasks. It outlines key components, real examples like PR babysitters and bug fixers, and explains why better models, built-in loop commands, and maturing toolchains make loops practical now.

• Opus 4.6 completes half of 12-hour tasks, 6x better than last year's model, and Stripe restructured its whole codebase in a day
• PostHog used a loop to fix a 3-year-old query-engine bug and got an 11% speedup
• Loops work now because of more stable models, built-in loop commands, and mature harnesses/toolchains (subagents, MCP, context compaction)
• The engineer's role shifts from writing code to defining goals, evaluation criteria, and context rather than being replaced

Nine copies of Claude Opus 4.6 were equipped with sandbox environments and tasked to autonomously develop weak-to-strong supervision methods, scoring their progress by “performance gap recovered” (PGR). The AARs reached a PGR of 0.97 versus a human baseline of 0.23, showed partial generalization to new tasks, but failed to replicate gains at production scale, underscoring both the promise and limits of automated alignment experiments.

• Nine Claude Opus 4.6 copies working autonomously hit a PGR of 0.97 on weak-to-strong supervision, crushing the human baseline of 0.23, in just 5 days and ~$18,000.
• The top method didn't generalize well: it worsened coding performance on unseen tasks and showed no significant gain when tested on Sonnet 4 with production infrastructure, suggesting it overfit to quirks of the original setup.
• Giving each AAR a different, even vague, starting prompt outperformed rigid, uniform workflows by avoiding convergence on the same dead-ends.

Researchers used a “concept injection” method to compare Claude’s self-reported thoughts with its actual neural activity. They found Claude Opus 4 and 4.1 sometimes detect and control injected concepts, suggesting limited but real introspective abilities that improve with model capacity.

• Anthropic injected known concept vectors into Claude's activations and found Opus 4/4.1 could sometimes notice and identify the "unexpected thought" before it appeared in output, suggesting real internal monitoring rather than post-hoc confabulation.
• Opus 4.1 only caught these injections about 20% of the time, and weaker models barely detected them at all, showing introspection is real but rare and scales with model capacity.
• Retroactively injecting "bread" into prior activations made Claude claim it had intended to say "bread" and fabricate a justifying backstory, indicating it was consulting an internal record of its own prior state rather than just rereading its text.
• Overly strong injections caused hallucinated, confabulated explanations (e.g., mistaking a "dust" vector for a literal speck), showing the effect is fragile and highly sensitive to injection strength.