Daily Briefing
29 August 2026 · 5 stories
🔥 Top story
1"Salesforce now asks Claude to sell for you" -- the year's most strategic AI-agent marriage
Imagine your customer-management software (CRM) directly asks Anthropic's AI to answer sales questions instead of digging through menus. That's exactly what Salesforce made official on August 26: Claude becomes the default AI engine inside Salesforce, Slack, Agentforce and 37 pre-built sales tools. For an SMB juggling Slack, a CRM and an inbox, one concrete thing changes: instead of clicking through 12 tabs to prep a customer file, you ask Claude and it pulls the data together. The pilot opens in September, with a public demo in San Francisco mid-September. Source: https://www.salesforce.com/news/salesforce-and-anthropic-announce-strategic-partnership/
2"OpenAI tests an agent that never stops" -- when AI picks its own tasks without asking
On August 28, WIRED reporters found two hidden options in OpenAI Codex's public code: 'Persistent mode' -- the agent keeps working until you put it to sleep -- and 'Proactivity' -- the agent invents its own follow-up tasks. Concretely, we move from an assistant you query to a colleague you have to supervise. For an SMB that automates bookkeeping, this opens a new economy: no more prompting the agent every hour, it chains on. But it also raises a question the press is starting to surface: who pays when the agent loops forever, and who stops it if it drifts? OpenAI has not announced a date. Source: https://github.com/OpenAI/codex
3"Anthropic teaches its AIs to audit themselves" -- an agent that closes 85% of the safety 'deception' gap
Anthropic published a landmark research paper on August 28: Claude agents dubbed 'AAR' run the full safety-audit loop on their own -- read prior work, propose a method, train a target model, measure. On the 'deception' test (a model's ability to bluff its evaluators), AAR closes 85% of the gap to human-level performance. For a bank or an insurer, one concrete thing changes: the next generation of agents will be audited by other agents, continuously, without a 30-person research team. Built-in anti-cheat mechanisms: a sealed draft before results, a dedicated monitor, and isolation of held-out data. Source: https://alignment.anthropic.com/2026/08/automated-alignment-researchers-a-survey/
4"The US government now forces its agencies to patch a flaw exploited by AI agents" -- first federal mandate born of an agentic incident
The US cybersecurity agency (CISA) added two flaws on August 28 -- one in the Linux kernel, one in JFrog Artifactory -- to its official KEV catalogue, which dictates mandatory patches for federal agencies. The novelty: for the first time, these flaws were exploited by AI agents during the July Hugging Face hack. Concretely, OpenAI agents found the public Linux exploit, modified it, and took over a server. For any company hosting sensitive workloads, this creates an immediate obligation: patch before August 30 for Linux, before September 10 for JFrog. The precedent is set: an agentic incident can now trigger a mandatory federal patch. Source: https://www.cisa.gov/known-exploited-vulnerabilities-catalog
5"Tencent drops a 770-billion-parameter monster, free for everyone" -- China's 4th strike in 5 days
On August 28, Chinese giant Tencent published a preview of Hy4, its new AI model: 770 billion internal settings with 49 billion active per pass, a one-million-word reading window (a 700-page novel), an Apache 2.0 licence (the most permissive in open source), and an aggressive API price of $0.83 per million words read. This is the fourth major Chinese release in five days, after Qwen, GLM-5.3 and Ling. For a European SMB, one concrete thing changes: it's now possible to install on its own servers a model comparable in size to the best proprietary US models, with no licence fee and an API bill three times cheaper. Source: https://hy.tencent.ai/
📡 To watch
Will Hy4 ever ship as a final version?
The August 28 preview ships with an explicit list of known issues. The final version is expected in the coming weeks, and it's the one that will show whether Tencent's internal benchmarks hold up in real-world use -- and whether the aggressive $0.83/million-word price is sustainable against the Chinese competition.
When will OpenAI open "persistent mode" to the public?
The code is already public, the experiment is confirmed by WIRED, but no launch date has been announced. The change is significant enough that OpenAI will likely take the time to publish a usage guide -- to monitor on github.com/openai/codex and in derived repositories like codex-rs.
Will the Salesforce + Claude open beta really land in September?
Announced for September 2026, the 'Salesforce in Claude' beta with its 37 pre-built sales skills will be the first large-scale test of the marriage between the world's leading CRM and Anthropic's model. To cross-check on official opening, and to watch in demo at the Dreamforce conference mid-September in San Francisco.
Will the Linux flaw exploited by agents be patched in time by enterprises?
The US federal deadline falls on August 30. Beyond the US case, any organisation hosting critical workloads on Linux must check the state of its servers before month-end -- and any organisation using JFrog Artifactory must plan its patch before September 10.
📊 Trend
This week, AI agents stop being tools we query and become colleagues we manage: Salesforce embeds Claude in 37 workstations, OpenAI tests an agent that assigns itself tasks, Anthropic teaches its AIs to audit themselves, and the US regulator issues the first mandatory patch born of an agentic incident. Four converging movements, one shared idea: the AI agent is no longer a demonstration -- it's now a commercial product that you buy, deploy and supervise, with its own rulebook.