Four operational tests displaced feature announcements. Codex exposed a rate-limit failure mode; domain-matched language evaded prompt-injection detectors; Ferrari attached an engagement claim to generated fan content; and Yugabyte showed the same Claude model changing behavior across skills, tool versions, and interfaces.
1. Codex rate-limit incident resolves after an overnight degradation
OpenAI reported an increase in users hitting Codex rate limits beginning at 4:37 p.m. on May 22. A mitigation entered monitoring later that evening, and the company marked all affected services fully recovered at 10:58 a.m. on May 23.
The status record supports a capacity incident rather than a deliberate pricing change or permanent quota cut. Queues, bounded retries, visible failure states, and a manual route can keep a throttled coding agent from silently repeating expensive requests on a critical delivery path.
Sources: OpenAI's Codex incident record
2. Domain camouflage defeats common prompt-injection detectors
A new preprint tests prompt injections rewritten to resemble the vocabulary and authority structure of their target documents. Across 45 tasks, detection reportedly fell from 93.8% to 9.7% for a Llama 3.1 8B setup and from 100% to 55.6% for Gemini 2.0 Flash; Llama Guard 3 detected none of the camouflaged payloads.
The sample is limited and the results await peer review, but the attack model is realistic: malicious instructions need not look like obvious overrides. Controls based on text classification alone remain probabilistic, so tool permissions, data isolation, output validation, and egress restrictions are necessary even when an injection detector is present.
Sources: Domain-camouflaged injection preprint, version 1
3. Ferrari and IBM use AI to reshape a year-round fan app
Ferrari's redesigned fan app uses IBM technology for AI-written race summaries, a question-answering companion, predictions, games, and analysis of engagement signals. Ferrari told TechCrunch that engagement rose 62% over race weekends after the partnership, a company-reported figure rather than an independently audited experiment.
The deployment is notable because the objective is sustained storytelling, not race engineering. Useful evaluation would connect personalization to retention and satisfaction while checking translation quality, source accuracy, consent for behavioral analysis, and whether generated summaries preserve the distinction between race data and editorial interpretation.
Sources: TechCrunch on Ferrari and IBM's fan platform · IBM's Scuderia Ferrari partnership hub
4. Coding-agent tests show context and harnesses can outweigh model labels
Yugabyte reported more than 350 evaluations across 17 model configurations for distributed-SQL coding tasks. Adding a YugabyteDB skill file raised anti-pattern avoidance from 2.42 to 3.79, but prose rules for retry control flow regressed; the same Claude model also varied substantially across tool versions and interfaces.
This is a vendor-designed benchmark for its own database, so absolute rankings are less useful than the failure analysis. Static prohibitions helped with product-specific facts, runnable examples worked better for procedures, and workload-specific tuning hurt broader tasks, supporting separate universal and project-level context plus execution tests on the real cluster.
Sources: Yugabyte's coding-agent benchmark report · YugabyteDB skill files