# AI Breaking Stories Edition 005 — Research Evidence

**Edition date under review:** 25 August 2026  
**Status:** Research evidence only; no publication authority.

## Candidate evidence

| Candidate | Source and date | Verified evidence | Required qualification |
|---|---|---|---|
| Open-weight model safeguards can be removed | University of Waterloo news release on EurekAlert, 25 August 2026 | An international team led by Waterloo and FAR.AI tested 21 popular open-weight LLMs using TamperBench and reported that all could be modified despite built-in safeguards. The release says the paper was presented at ACM KDD and links DOI `10.1145/3770855.381755`. | This is an institutional news release about a research paper. The authors state that weaknesses may not be unique to open models and that openness remains important for research and accountability. |
| Agent adoption is rising faster than operational assurance | Temporal State of Development Report 2026, 25 August 2026 | Temporal reports 554 retained respondents from a Qualtrics survey of current AI-agent users in the US and UK/EMEA. It reports 80.8% daily-or-more use, 49.1% saying agents are in production or core to shipping, 41.1% encountering issues daily or more, state tracking as the top blocker, and 39.5% citing security concerns as a blocker to self-managing agents. | Vendor-sponsored survey of a selected cohort already using agents. “Successful” status and many performance measures are self-reported; the findings are not representative of all engineers or organisations. |
| AI is being embedded into operating-model design | Kyndryl and Bridgepath press release, 25 August 2026 | The companies announced a collaboration to design, build, operate and scale AI-native global capability centres, explicitly embedding AI into delivery, operations, decision-making and governance with defined service structures and in-house capability. | Productivity and business-outcome statements are company assertions for a proposed service model, not independent outcome evidence. |
| Construction risk support is being tested with evidence and rationale | Fujitsu, Tokyu Construction and Kitano Construction press release, 25 August 2026 | The companies announced an August–December field trial analysing schedules, reports, procedures, inspections and other site data to identify possible omissions and delay risks one to two months ahead, with supporting rationale for site managers. The trial will assess accuracy, usefulness and workload impact. | This is a field trial, not a proven deployment result. The system supports site managers; the announcement does not transfer construction or safety authority to the AI. |
| Organisational readiness is being reframed as orchestration | Open-access *Discover Artificial Intelligence* study, published 25 August 2026 | The qualitative study combines a literature-derived analytical template with 15 semi-structured interviews across corporate and SME contexts. It proposes Organisational AI Orchestration Capability and reports governance/accountability, process discipline and behavioural alignment as primary deployment determinants once infrastructure and data become baseline prerequisites. | Exploratory qualitative research with a small practitioner sample. It develops a conceptual framework rather than proving causal performance effects. |

## Sources

1. https://www.eurekalert.org/news-releases/1141253
2. https://temporal.io/reports/state-of-development-2026
3. https://www.kyndryl.com/us/en/about-us/news/2026/08/ai-native-global-capability-centers
4. https://global.fujitsu/en-001/pr/news/2026/08/25-01
5. https://arxiv.org/abs/2602.06911
6. https://link.springer.com/article/10.1007/s44163-026-02019-z

## Screened but excluded

| Candidate | Decision | Reason |
|---|---|---|
| AI-initiated commerce and payment authentication | Exclude from Edition 005 | The strongest accessible industry article, PYMNTS’ *Banks Prepare Payments for AI Commerce*, is dated 21 August 2026 rather than 25 August. It may be relevant to a future thematic analysis, but it is not new enough for this dated edition. |
| AI infrastructure financing | Exclude from Edition 005 | Infrastructure financing was substantively covered in Edition 003. No independently verified 25 August development justified repetition. |
| Kyndryl–Bridgepath AI-native GCC collaboration | Exclude from the four-story draft | The 25 August announcement is verified, but its operating-model message overlaps the stronger open-access organisational-readiness study published the same day. The academic study offers clearer methodology and explicit SME relevance. |

Excluded commerce source reviewed: https://www.pymnts.com/news/payments-innovation/2026/banks-prepare-payments-for-ai-commerce/ (published 21 August 2026).

## Publication boundary

These findings are candidates for a review draft only. Edition 005 remains absent from the live registry, sitemap and metadata until Phillip Llewellyn explicitly authorises the completed edition.