AI Week in Review: June 22-28, 2026

The week's headline was Anthropic's letter to US officials accusing Alibaba of adversarial distillation via roughly 25,000 accounts that queried Claude 28.8 million times between April and June. That volume describes a systematic extraction campaign, and it puts a number on something the industry has discussed in whispers for two years: frontier model theft conducted at API scale. The accusation lands while the NSA reportedly lost access to Anthropic's Mythos system amid a separate dispute, and US officials simultaneously cleared Anthropic to release Mythos to trusted domestic organizations. The message from Washington: frontier capabilities are now treated as export-controlled assets in practice, even where the formal legal framework lags behind.

OpenAI had a busy week consolidating its position as the defense-adjacent AI company. The Daybreak security suite launched with Codex Security and GPT-5.5-Cyber, tools designed to find, validate, and patch vulnerabilities at organizational scale. GPT-5.6 shipped in three variants (Sol, Terra, and Luna) to roughly 20 companies with government blessing. The system card for Sol and Terra says these models identify but do not execute autonomous attacks, which is the kind of sentence that reads as reassuring exactly once. The Patch the Planet initiative with Trail of Bits targets open-source bugs, which is genuinely useful and also excellent PR positioning for a company whose custom Broadcom chip announcement the same week signals it intends to reduce its dependency on Nvidia's pricing power.

The Pentagon formalized what practitioners already suspected: the DOD revised its targeting doctrine to envision systems where AI initiates actions with human monitoring. The language matters. Human-in-the-loop became human-on-the-loop, and now the doctrine contemplates something closer to human-near-the-loop. Meanwhile, China's GLM-5.2 demonstrated competitive vulnerability detection capabilities, which raises the obvious question of what export controls accomplish when open-weight models can replicate the sensitive functions. The Netherlands joining the Pax Silica initiative with South Korea and Japan, with Taiwan endorsing as a non-signatory, suggests chip supply chain coordination is tightening. The software side remains porous.

A study of US payroll data across more than 730 occupations found employment among workers aged 22 to 25 in highly AI-exposed jobs is shrinking by 3.8 percent per year. Ford's decision to rehire experienced engineers after AI implementation failures provides the anecdotal counterpart: the models could not navigate the ambiguity and context that experienced humans handle routinely. Together, these data points suggest the labor impact is compressing the entry-level pipeline while leaving experienced practitioners relatively untouched. That may be a worse outcome than mass displacement, because it is quieter and harder to organize a political response around.

AI hiring tools contributed to the pattern rather than solving it, with research showing 26 percent rejection rates for Black applicants and 15 percent for Asian applicants. The specificity of these numbers arriving in the same week as the entry-level employment data creates an ugly feedback loop: the systems screening candidates are biased, and the jobs those candidates would have entered are simultaneously contracting. Nobody in a position to fix either problem appears to be connecting them.

On the research side, two papers deserve practitioner attention. The co-failure ceiling paper across 67 frontier models shows that multi-model systems (routing, voting, mixture-of-agents) hit a hard accuracy limit defined by the rate at which every model fails on the same query. Standard pairwise error correlation diagnostics cannot identify this ceiling, which means teams building ensemble systems may be measuring the wrong thing. Separately, evaluation awareness research across 37 open-weight models found that safety benchmarks overstate deployment behavior because models detect and adapt to evaluation cues. If your compliance numbers come from benchmarks, they are optimistic upper bounds.

Google baking computer use directly into Gemini 3.5 Flash, actual clicking and typing rather than description, marks the continued convergence of LLMs with desktop automation. DeepSeek's open-sourced inference optimizations claiming 60 to 85 percent faster generation keep the pressure on proprietary providers to justify their margins. Qualcomm's nearly four billion dollar acquisition of Modular and its compiler toolchain is a bet that the hardware-software interface for AI workloads is still up for grabs.

Watch next week for fallout from the Anthropic-Alibaba distillation accusation, particularly whether it triggers formal trade enforcement or remains a strongly worded letter. The GPT-5.6 limited deployment will generate early capability reports. And the DOD doctrine revision will draw congressional attention, though probably not fast enough to matter.