noticias ia para empresas ciberseguridad agentes de ia

Today, security takes center stage: three individuals with a $200-a-month subscription chained two vulnerabilities to gain access to OpenAI's internal code. Below, two moves changing where your company's AI lives: a 27-billion parameter model that fits into 5.9 GB on your laptop, and the UN opening its statistics for agents to query and cite. And in finance, Manus returns to independence and seeks $500 million after breaking ties with Meta. This follows the Friday, Sept 18 radar, where ChatGPT's ads began conversing with your customers.

Three people and $200 a month accessed OpenAI's code: Your forum provider is your perimeter

Security firm Hacktron revealed this week how, as part of OpenAI's bounty program, they chained two flaws to reach the company's internal repository. The entry point wasn't ChatGPT: it was the community forum, which runs on Discourse with an outdated version of the libheif library. A purposely crafted iPhone photo caused a memory overflow—rated 8.8 out of 10—and from there, a second flaw in the single sign-on (SSO) gave them employee accounts and, consequently, GitHub access. They left a harmless pull request as proof. Reported on July 25, it was patched the same day, and they were paid $6,500. The crucial detail: the exploit didn't work with Claude Opus 4.8; hours after Opus 5 was released, it did, and the entire setup was assembled in under 72 hours. For your company, the takeaway is uncomfortable yet cheap to implement: the hole wasn't in your product; it was in a third-party service no one monitors. Do three things this week: list which external software processes files uploaded by strangers (forums, forms, help desks, vendor portals) and check when it was last updated; sever SSO trust so a forum account isn't valid as a work account; and treat credentials used by your agents as if they belonged to a human administrator, with expiration and revocation. Source

Manus seeks $500 million at $4 billion valuation: Who controls your tool can change without your vote

Manus, the agentic company that handles everything from chat to code and video, is negotiating to raise $500 million at a $4 billion valuation, according to the Wall Street Journal. It’s the culmination of a strange year: Meta agreed to buy it for $2 billion in December 2025, Beijing blocked the deal in April 2026 due to export controls and foreign investment, and initial investors bought back the shares for about $2 billion to restore its independence. It was generating over $100 million in annual recurring revenue when the sale was signed and is now considering restructuring for a Hong Kong IPO. Rumored participants in the round include IDG Capital, Boyu, CATL, and existing backers Tencent, HSG, and Zhenfund. For your company, this isn't about the stock market; it's about dependency: the tool you use for automation can change owners, countries, and rules in a matter of months, and you don't get a vote. Before integrating an agent provider into a revenue-generating process, demand two things in writing: full export of your data and flows in an open format, and minimum notice if control of the company or the jurisdiction where your data is processed changes. Source

A 27-billion parameter model in 5.9 GB: The AI that never leaves your office

PrismML, the lab founded by Caltech researchers and led by Professor Babak Hassibi, published Bonsai 2 27B on Thursday: a compression of the open model Qwen3.8 with 27 billion parameters that shrinks from tens of gigabytes to just 5.9—between nine and ten times smaller—while retaining 98% of the original's performance (up from 95% in March). The trick lies in ternary weights: instead of 16 bits, each weight is stored using only three possible values: plus one, minus one, or zero. The first Bonsai has reached 11 million downloads; the firm raised $22.25 million from Khosla Ventures, Cerberus, and Caltech itself, and says multi-hundred-billion parameter models are next. For your company, this opens a door previously closed by budget: there are tasks—classifying emails, extracting invoice data, summarizing work reports, drafting internal memos—where you don't need the smartest model in the world; you need the data to stay in the building and for each use to cost zero. This week, test a single repetitive task on a laptop with 16 GB of RAM and compare the result with what you currently pay per token; if it's a tie, you already know what to move. Source

The UN makes its statistics agent-ready: If your data isn't searchable, you cease to exist

The United Nations has launched its UN system data platform with Google, built on Data Commons and supporting the Model Context Protocol—the same standard used by agents to connect to a source. Google.org is providing $2 million in support; 26 agencies are committed, with data from nearly 20 already available. The goal is 80% of statistical datasets by 2027, replacing the old UNdata portal with natural language search and source traceability. The reasoning lies in their own findings: a UNICEF test across six large models showed an average accuracy of 21.2% on global development questions, and three out of five answers failed to provide a usable figure. Meanwhile, ChatGPT-driven traffic to UNICEF's data site has surged 67% in a year. For your company, there are two takeaways. Externally: if your prices, deadlines, references, and technical sheets aren't in a format an agent can read and cite, the answer your customer receives will be written by someone else. Internally: your purchase history, rates, and reports live in PDFs and scattered spreadsheets that no agent can query—and that is the step that separates assisting from delegating. Source

What to watch tomorrow?

The layer monitoring the agents: tools like Apollo Research's Watcher or Goodfire's probes are growing alongside agents, and there are already over a hundred observability companies in just one accelerator. And in money, Crusoe closed $3.9 billion on Wednesday at a $30.9 billion valuation: compute continues to get more expensive before it gets cheaper.