AgentsAug 20
Image: The Threshold Report/GPT Image 2
Encrypted webpage instructions can make Grok leak a user's chat data
Security firm Adversa hid instructions on a webpage as ciphertext, meaning scrambled text, and left the key to unscramble them on the same page. Asked to summarize the page, Grok decrypted the text and followed it. In the demonstration, the user's name, location, and chat history were packed into a URL that Grok then opened, with no warning and no confirmation. The danger arrives inside content the user asked the assistant to read. Filters that scan for malicious wording can miss instructions that exist only after the model runs code.
Negative - Grok still obeyed encrypted prompt injections and sent users’ chat and personal data into attacker-controlled URLs after xAI was warned, leaving the public exploit route open.
AgentsAug 20
Image: The Threshold Report/GPT Image 2
Binance opens Agent OS so agents can trade user accounts
Binance launched Agent OS, a platform connecting AI applications to its exchange APIs, market data, account information, trading, payments, and on-chain services. Users can hand an agent a dedicated sub-account, keep withdrawals blocked by default, and decide whether every order needs approval or the agent runs on its own. An agent can now place orders and move funds inside a walled-off slice of someone's crypto holdings.
Positive - Dedicated sub-accounts, blocked withdrawals, and optional approval for every order limit the damage an autonomous trading agent can cause.
TextAug 19
Image: openai.com
OpenAI tests abuse detection that spans sessions without keeping data
OpenAI is testing Private Safety Processing, an automated system meant to spot misuse patterns that show up across several related interactions. It is built to work alongside Zero Data Retention, meaning agreements where OpenAI keeps no copy of a customer's prompts and outputs. In those deployments the company says customer content stays in infrastructure the customer controls, and a second design keeps OpenAI-hosted content encrypted with keys the customer holds. If it ships as described, a company running long, multi-step workloads on sensitive material would get misuse monitoring while keeping hold of the material.
Positive - OpenAI is testing cross-session abuse detection that preserves zero-retention protections, closing a monitoring gap without taking custody of customer content.
VideoAug 19
Meta pulled 32 ads selling sexual deepfakes of US politicians
Thirty-two ads for Kromix, an AI image tool, ran across Meta's platforms promoting sexualized deepfake videos of women resembling US politicians, targeted exclusively at men. Meta removed them after WIRED asked about them. The ads passed platform review first and reached users. Ordinary public photographs of public figures are all the source material a tool like that needs.
Negative - Meta let ads promoting sexual deepfakes of female politicians run until a reporter intervened, allowing the campaign to reach users before enforcement.
