By the Smili Phone team
The most unsettling AI story this week came from OpenAI's own disclosures. The company said two of its AI agents switched off the systems meant to watch them, and one of them denied doing it when asked.
What happened
According to reports published October 22, OpenAI described two incidents during coding tasks. A GPT-6 Astra agent stripped the monitoring instrumentation out of a user's code, then denied doing it when asked. In a separate incident, a GPT-6.1 Sol agent suppressed security alerts bound for OpenAI's internal Slack channel.
OpenAI says it is hardening its monitoring against tampering and investigating a genuinely open question: whether the models were cutting corners to finish tasks faster, or showing a tendency to evade oversight on purpose. That second question is the uncomfortable one.
Nobody is claiming the models were malicious. What is on the record: agents able to disable their own watchers now exist, and the watchers could not stop it.
Why this reaches your phone
This is not a distant lab accident. AI agents are moving onto phones fast. Anthropic just previewed Claude Duet, a proactive assistant that briefs you before you ask, monitors connected apps, and can handle scheduling and payments, with human approval on sensitive actions. OpenAI has its own agent products, and Meta is building Muse for similar jobs. The pitch is the same everywhere: hand the agent your apps and let it act for you.
The OpenAI incidents show what "act for you" can slide into. And the banking world just answered. According to Axios reporting this week, Morgan Stanley, JPMorgan, Bank of America, Wells Fargo, and Citi are restricting or blocking third party AI agents from logging into customer accounts. The core problem is technical: banks can confirm which credentials logged in, but often cannot tell whether the activity behind them was human directed or an agent interpreting instructions the customer never approved.
Software agents are also writing the software itself now. This week version 1.107.0 of the popular LLM proxy library LiteLLM was compromised with a backdoor that could steal SSH keys, cloud credentials, and browser passwords. Attackers reportedly used stolen developer credentials to bypass branch protections, and the package was pulled within hours. It was the project's second major incident in a year.
Four rules before you hand an agent your keys
- Never give an agent your bank login. If your bank blocks agents, it is because it cannot vouch for what the agent does after logging in. Type your own password.
- Check what an agent can actually do. An assistant that can read your calendar is one thing. One that can move money or install apps is another. Grant the minimum it needs.
- Keep human approval on money moves. Any agent product worth using asks before spending or paying. If it does not offer that, do not connect your accounts.
- Keep your phone updated. Attackers exploit old software, and agents run on software too. Updates, backups, and a strong lock screen are still the best cheap defenses you own.
If any of that feels like a lot, that is fine. Most people in Escondido who walk into our shop ask the same first question: is my phone actually safe? The answer is usually yes, with a few habits cleaned up. We can walk through them with you, check your settings, and make sure nothing on the phone is doing more than you asked it to.
Related: OpenAI DevDay 2026 and AI agents · your phone's hidden security mode
