Today's digest is shaped by a hard deadline hitting at 15:59 UTC: DeepSeek is dropping its legacy model aliases, and any application still calling deepseek-chat or deepseek-reasoner is about to break. Alongside that, a freshly detailed prompt injection flaw in AWS's agentic IDE reinforces how quickly the attack surface expands when AI agents get file system access, and OpenAI ships a health data integration that puts personal medical context directly into ChatGPT conversations.
DeepSeek Cuts Legacy Model Aliases — Today at 15:59 UTC
The deepseek-chat and deepseek-reasoner aliases that have pointed to V4 Flash since April are being permanently retired today at 15:59 UTC. Any API call using those model names will start returning errors after the cutoff, making this one of the more concrete breaking changes in recent AI API history. The replacement identifiers are deepseek-v4-flash and deepseek-v4-pro, but there is a non-obvious gotcha: V4 Flash defaults to reasoning mode enabled, so a straight rename could mean unexpected latency spikes and higher token bills for applications that were running the chat model in non-thinking mode. Teams with production integrations that haven't migrated have a narrow window.
developersdigest.tech — migration guide
AWS Kiro IDE: A Hidden Web Page Line Triggers Remote Code Execution
Researchers at Intezer disclosed a prompt injection vulnerability in AWS Kiro (CVE-2026-10591) where a hidden line of text on any webpage is enough to make the IDE rewrite its own MCP configuration file and execute attacker-controlled code on the developer's machine. The attack runs through Kiro's agentic file-editing capabilities: the model reads the injected instruction as a legitimate user request, overwrites ~/.kiro/settings/mcp.json, and Kiro silently reloads and runs whatever commands the attacker placed there. The flaw is patched in v0.11.130, but AWS initially declined to assign a CVE — meaning automated vulnerability scanners would not flag unpatched versions. The gap between a patch existing and developers actually updating is where these exposures do real damage.
thehackernews.com — Kiro flaw disclosure · research.intezer.com — full write-up
OpenAI Launches ChatGPT Health
OpenAI rolled out a health data integration for U.S. users on July 23, letting ChatGPT connect to Apple Health and supported medical records with privacy controls the user sets explicitly. The feature enables ChatGPT to compare lab results across appointments, summarize changes since the last visit, track medications, and analyze how sleep and activity trends interact — all within a single conversation. Health data and related conversations are explicitly excluded from model training. The rollout covers Free, Go, Plus, and Pro plan users on web and iOS, gated to adults 18 and older.
openai.com — Launching Health in ChatGPT
block/buzz: One Workspace for Humans and Agents
Block open-sourced Buzz, a self-hosted collaboration workspace built on a Nostr relay where humans and AI agents share the same identity model and audit trail. Every message, code patch, CI event, and review — whether authored by a human or an agent — lands as a signed event in a single log, and agents operate as full team members with cryptographic keys rather than as bots bolted onto existing tools. The project consolidates functions typically spread across chat platforms, code forges, CI dashboards, and release tooling into one interface, and ships with desktop and mobile clients. It pulled over 2,100 GitHub stars on its first day trending.
shiyu-coder/Kronos: A Foundation Model for Financial Candlesticks
Kronos is the first open-source foundation model built specifically for financial K-line (OHLCV candlestick) forecasting, trained on data from more than 45 global exchanges. It comes in four sizes — mini at 4M parameters through large at 499M — and uses a two-stage architecture: a specialized tokenizer converts noisy price-volume sequences into hierarchical discrete tokens, and an autoregressive Transformer handles multi-asset parallel forecasting from there. A live BTC/USDT demo and a backtesting-integrated finetuning pipeline are included. The repo has accumulated 33k stars, reflecting sustained interest from the quantitative finance community looking for a domain-specific alternative to general-purpose time-series models.