AI News Roundup: October Twenty-Five - The Audit Turn

Overview: Today the industry stopped arguing about capability and started arguing about proof. Apollo's chief economist published numbers showing AI has produced no measurable payoff for most of corporate America, Anthropic told builders to delete most of their agent architecture, and a model-bias study showed two chatbots recommending opposite candidates in nearly every midterm race. Add nations hedging between the US and Chinese AI blocs and a leaky PHP disclosure embargo, and the thread is clear: everyone is auditing what AI actually does, not what it promises.

---

Grok and GPT pick opposite candidates in nearly every race

A study of nearly 500 US midterm races by Convovo found Grok recommending Republicans in 92% of races and GPT recommending Democrats in 97%. The split is not explained by the base weights alone: the divergence lines up with retrieval, meaning the search layer is doing as much political work as the model itself, according to TechCrunch.

For builders this is the uncomfortable part. If you ship an agent on top of a third-party API, your system prompt and your retrieval choices can push answers toward conclusions you never wrote. The practical test is cheap: run the same eval with search on and with search off, and log the tokens you send against the tokens you are billed for.

---

Kazakhstan shows how nations hedge between US and China AI blocs

Washington and Beijing are building rival AI stacks, complete with export controls and procurement rules, and middle powers are refusing to pick a side. Kazakhstan and Indonesia are signing with both, treating AI suppliers the way they treat any other strategic input, as AP reported.

The consequence lands on engineering, not diplomacy. If customers can sit in either bloc, the inference layer has to route across American frontier APIs and Chinese open-weight models without the product knowing the difference. Export licensing becomes an architecture constraint rather than a legal afterthought.

---

Anthropic tells builders to fold subagents into one agent loop

Anthropic published guidance arguing that a single agent loop with skills now beats specialised subagents on quality, cost and latency, and shipped a migration kit to do the folding, per its engineering writeup.

This is the quiet reversal of a two-year trend. Most multi-agent harnesses exist to compensate for context limits and instruction-following failures that newer models no longer have, which means a lot of production routers and specialist crews are now overhead. Keep subagents where they buy real parallelism or contain genuinely separate context. Fold everything else, but do it against an eval set rather than a vibe.

---

AI found PHP's holes first. Then the secret leaked

AI-assisted research surfaced critical vulnerabilities in PHP, the runtime under a large share of the web, and governments responded by taking systems offline. The messier story is what happened next: the coordinated disclosure embargo cracked before patches were broadly deployed, per CISA's emergency alert and BBC coverage.

Discovery is now faster than distribution. Any team with a slow deploy pipeline had an exploitable window measured in hours, and the same logic will apply to the next vulnerability class AI finds. Inventory your runtimes, including the ones hiding inside WordPress installs, vendor appliances and sidecars, and pre-stage mitigations you can flip without a full release.

---

Apollo's chief economist: AI delivers zero financial payoff outside tech

Torsten Slok of Apollo Global Management looked at the S&P 493, the index minus Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla, and found flat profit margins and flat revenue per worker. His conclusion, reported by TechCrunch, is blunt: the technology has delivered zero financial payoff to corporate America outside of tech.

Slok's explanation is worth taking seriously rather than dismissing. Cheap intelligence is not the same as judgment, and judgment carries accountability, trust, culture and unwritten rules. That is the gap automation projects keep hitting: the work that gets delegated is the work that was already documented, while the value sat in the parts nobody wrote down.

---

What to Watch Tomorrow

• Grok's AI-generated film 'Odyssey' released its first seven minutes - a generated feature is a distribution and rights story before it is a quality story, via TechCrunch.
• AI wearables and gadgets - the hardware category keeps consolidating around a handful of form factors, per TechCrunch.
• Fallout from the PHP vulnerabilities - watch disclosure policy, not patch notes. The coordination model is the thing under strain.

Builder Impact

• Prove the payoff or lose the budget. Slok's flat S&P 493 numbers will be quoted in every CFO conversation this quarter. Ship one metric tied to a process you actually changed, not to tokens consumed.
• Simplify before you scale. Anthropic's guidance is a permission slip to delete architecture. Run an eval, fold the agents that chain on each other, and keep the ones that isolate context.
• Audit the retrieval layer, not just the model. The Grok and GPT split shows search can move outputs more than weights do. Test both modes.
• Treat patch latency as your real security posture. AI finds holes faster than organisations deploy fixes, and that ratio is the exposure measure that matters.
• Design for a two-bloc world. Routing between US frontier APIs and Chinese open-weight models is now a product requirement for anyone selling internationally.

The day's pattern is not that AI failed. It is that the free-credit phase of AI adoption ended: capability claims now need audit trails, and the teams that ship them will keep their budgets.