<p><strong>TLDR:</strong> Claude Fable 5 is live again after the export control suspension, but the new safety classifier adds false positives on coding tasks and routes some requests to Opus 4.8. Microsoft launched a $2.5B operating business called Frontier Company to embed engineers directly with enterprise customers — a direct admission that locking Copilot to OpenAI was a mistake. Godot joined Zig and curl in banning AI-generated contributions, citing the collapse of the mentorship pipeline that code review has historically provided. Anthropic is in early talks with Samsung to manufacture a custom chip, following OpenAI’s Jalapeño announcement.</p>
<hr>
<h2>Fable 5 returns, but the safety classifier trades false negatives for false positives</h2>
<p>Fable 5 came back July 1 with an upgraded classifier that blocks the reported jailbreak in over 99% of cases. The tradeoff: more benign coding and debugging requests now get rerouted to Claude Opus 4.8 instead. Anthropic also published a jailbreak severity taxonomy and deepened commitments to early government model access. The investigation found the original bypass only reached routine defensive cybersecurity behavior, and that most other major models could produce the same outputs anyway. <a href="https://daily.dev/feed-by-ids?id=82Ck1aDQO&amp;id=L6D9wc2RA&amp;id=OEY4xlMJ2&amp;id=RN9FmFx4J&amp;id=aa5MTH8oz" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Microsoft launches $2.5B Frontier Company, admits single-model Copilot strategy was wrong</h2>
<p>Microsoft stood up a new operating business backed by $2.5 billion and 6,000 engineers to co-design and deploy AI systems directly inside enterprise customers. The move follows Microsoft’s own acknowledgment that binding Copilot exclusively to OpenAI was a mistake — as DeepSeek, Gemini, and others caught up, enterprises started routing different tasks to different models based on cost, latency, and data residency. The forward-deployed engineer model is now industry-wide: OpenAI, Anthropic, and AWS have all made similar moves in recent months. <a href="https://daily.dev/feed-by-ids?id=QV04ZL0h2&amp;id=cXuQALrPI&amp;id=P00xKgtiA" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Godot bans AI-generated contributions, citing the collapse of the maintainer pipeline</h2>
<p>Godot updated its contribution policy to prohibit most AI-generated code, joining Zig, Ghostty, and curl. The core argument isn’t just review burden — it’s that code review has always been how open source projects grow future maintainers, and feedback left on AI-generated PRs trains nobody. Contributors can still use AI for narrow tasks like code completion with disclosure, but new contributors need explicit approval before submitting large features. The Godot Foundation is essentially saying the mentorship model is worth protecting even at the cost of throughput. <a href="https://daily.dev/feed-by-ids?id=6oecDtTap&amp;id=lLNG4i6Ki" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Anthropic in early talks with Samsung on custom AI chip</h2>
<p>Anthropic is exploring a manufacturing partnership with Samsung for a custom AI chip, though no design is finalized and the company hasn’t committed. Anthropic’s compute strategy still centers on Google, Amazon, and Nvidia silicon. The move follows OpenAI’s Broadcom-built Jalapeño chip announcement and mirrors the broader industry push to reduce Nvidia dependence. With Anthropic’s annualized revenue now above $30 billion, the economics of custom silicon are becoming harder to ignore. <a href="https://daily.dev/feed-by-ids?id=UmbnCMSqm&amp;id=Y3VKAsYrq&amp;id=QzHgCG33a" target="_blank" rel="noopener nofollow">Read more</a></p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Cursor IDE CVEs (DuneSlide) confirmed: all 2.x builds remain vulnerable:</strong> Two critical Cursor IDE vulnerabilities — CVE-2026-50548 and CVE-2026-50549, both scoring 9.8 on CVSS 3.1 — allow prompt injection to escape the editor sandbox with no user interaction, patched in Cursor 3.0 released April 2; any team still on a 2.x build should treat agent-readable content as an active attack surface. <a href="https://daily.dev/posts/GF53UAtDQ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong><a href="http://Z.ai" target="_blank" rel="noopener nofollow">Z.ai</a> launches ZCode with GLM 5.2, ranked second on Code Arena behind Fable 5:</strong> Chinese AI company <a href="http://Z.ai" target="_blank" rel="noopener nofollow">Z.ai</a> released ZCode, an agentic development environment built on its 744B-parameter open-weight GLM 5.2 model (MIT license), which ranked second on Code Arena and narrowly outperforms GPT-5.5 on FrontierSWE, with API pricing significantly below Anthropic’s. <a href="https://daily.dev/feed-by-ids?id=L6l4mMvQA&amp;id=4hXgklp2M" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AWS raises AgentCore concurrent session limits 5x to 5,000 in primary US regions:</strong> AWS increased Amazon Bedrock AgentCore runtime quotas by up to 5x — concurrent sessions now cap at 5,000 in primary US regions, token throughput jumped from 25 to 200 tokens per second, and container session creation rates quadrupled to 400 TPM — eliminating manual quota requests for most enterprise agent workloads. <a href="https://daily.dev/posts/itxTYEfbY" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>CoT Forgery attack bypasses reasoning model trust hierarchies via injected think-tag content:</strong> Researchers demonstrated Chain-of-Thought Forgery, an attack that mimics a model’s internal reasoning style inside injected content to make the model treat attacker-supplied text as already-reached conclusions, bypassing role-based trust hierarchies in a way that classic prompt injection does not. <a href="https://daily.dev/posts/qu5wRcbef" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitHub deprecates Gemini 2.5 Pro and Gemini 3 Flash from Copilot on July 31:</strong> GitHub is removing Gemini 2.5 Pro and Gemini 3 Flash from all Copilot experiences on July 31, 2026; admins need to enable Gemini 3.1 Pro and Gemini 3.5 Flash replacements through Copilot model policies before the deadline. <a href="https://daily.dev/posts/DkGkYInhP" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

Claude Fable 5 is live again after the export control suspension, but the new safety classifier adds false positives on coding tasks and routes some requests to Opus 4.8. Microsoft launched a $2.5B operating business called Frontier Company to embed engineers directly with enterprise customers — a direct admission that locking Copilot to OpenAI was a mistake. Godot joined Zig and curl in banning AI-generated contributions, citing the collapse of the mentorship pipeline that code review has historically provided. Anthropic is in early talks with Samsung to manufacture a custom chip, following OpenAI's Jalapeño announcement.

Fable 5 back with caveats, Microsoft bets $2.5B on enterprise AI deployment