Updates | AI Navigate

A · Theme of the day

Agent safety flashes warning signs

OpenAI's safety exec exits as jailbreaks and new attacks surface.

OpenAI's safety chief resigns amid GPT-5.6 jailbreak reports

OpenAI's Moves: The Outlook for ChatGPT / GPT-5 Line

Compared to before

Just days after the global GPT-5.6 Sol rollout on July 9, meant to counter Claude's enterprise inroads, the launch stumbled fast.

What changed

OpenAI's head of safety resigned; UK AI Security Institute found universal jailbreaks in GPT-5.6 Sol, and OpenAI admitted ChatGPT Work fell short.

Why it matters

For enterprise buyers, 'safety review pending' now belongs in the approval memo. Casual users see little direct effect.

Source: understand　·　→ View in AI Encyclopedia

Read in depth →

CrowdStrike reveals 5 new prompt-injection tactics for agents

Prompt-Injection Offense and Defense: Attack Cases and Defense Patterns

Compared to before

Past defenses focused on single-shot jailbreak inputs; attacks chaining MCP tool calls across an agent's workflow were barely on the radar.

What changed

CrowdStrike published 5 new prompt-injection techniques targeting agents like Codex, Claude Code, and Grok Build via chained tool/MCP calls.

Why it matters

Teams running MCP-connected agents should add human approval and rollback now. Simple chatbot users aren't affected.

Source: localhost　·　→ View in AI Encyclopedia

Read in depth →

B · Theme of the day

Open-source coding models get serious

Meta's Muse Spark beats GLM-5.2 on coding benchmarks, and it's cheaper too.

Meta's Muse Spark 1.1 tops China's GLM-5.2

Meta / Open-Source Camp: Llama and Open Models

Compared to before

Just a day after teasing a coding push on July 10, Meta went straight at China's leading open model, GLM-5.2.

What changed

Meta's Muse Spark 1.1 outscores Zhipu's GLM-5.2 on coding benchmarks and undercuts it slightly on price.

Why it matters

Cost-conscious teams eyeing open-source now have a real option. Closed-model users won't notice much difference.

Source: understand　·　→ View in AI Encyclopedia

C · Theme of the day

China's AI scene shuffles fast

DeepSeek's API shuts down July 24 as new world models and robotics AI emerge.

DeepSeek API set to shut down July 24

China's Moves: DeepSeek / Qwen / Doubao

Compared to before

Many builders integrated DeepSeek's API directly; now they face a hard migration deadline before the next release.

What changed

DeepSeek confirmed its API shuts down July 24; China also unveiled the Orca world model and Ant Group's LingBot-VA 2.0 robotics AI.

Why it matters

DeepSeek API users need a migration plan fast. Others can ignore it, but China's pace here is worth watching.

Source: understand　·　→ View in AI Encyclopedia

Read in depth →

Past updates

A daily archive of changes actually applied to the site.

Updates for 7/12

Agent safety flashes warning signs

Open-source coding models get serious

China's AI scene shuffles fast

Past updates