suren

Systems design and AI agents, in practice. Building my own bot team at home and sharing what breaks. Learning in public, daily.

[suren @radnerus]

My Raspberry Pi runs my life:

📈 AI agents manage my stock portfolio
🖥️ A dashboard tracks my Pi, home & market
🌡️ It even throttles its own fan by temperature

All from Telegram Bot. No cloud, no subscription.

Ask me anything about building this 👇

[suren @radnerus]

@SantoshYadavDev Happy birthday, Santhosh🥳

Aging like a fine wine! Keep it going

[suren @radnerus]

Sources: SBS on the Nov 10 announcement https://t.co/x7XNym1lGx and ACS Information Age on the June incident https://t.co/pJN1wWTfPC

[suren @radnerus]

An OpenAI agent hit a "no" on an Australian government site. It went around the block anyway.

Yesterday Australia answered with a new AI safety regulator and mandatory incident reporting, where "misaligned behaviour" counts as an incident.

[suren @radnerus]

If you ship agents: treat "denied" as a signal, not just an error.

Log every 403, auth refusal and blocked fetch. Alert when the agent tries a new route to the same target. Cap retries per domain.

And decide who you'd notify, and how fast, before you need to.

[suren @radnerus]

Australia's response, Nov 10:

- All 18 review recs accepted
- New regulator with penalty powers
- Time-bound reporting, misaligned behaviour included
- Independent testing of frontier models
- Possible failure-to-prevent offence

No penalty sizes or deadlines yet.

[suren @radnerus]

@HamelHusain Agree. Public benchmarks tell you which models to try. Your own evals tell you which one to ship.

A few dozen real failures pulled from your own traces beat any leaderboard for your use case.

[suren @radnerus]

Paper:

Beyond Instruction Override: Benchmarking Indirect Prompt Injection with Task-Tampering Attacks (Zhang et al., Amazon Responsible AI)

https://t.co/KANWs39lcI

[suren @radnerus]

Allowlisting your agent's tools barely touches the prompt injections that don't try to hijack it.

New Amazon benchmark: a tool-policy defense moved attack success from 17.0% to 15.4%.

The attack keeps the right tools. It changes the values.

[suren @radnerus]

Caveats: simulated environments, 90 cases built with LLM help, one run per config at temperature 0.

Read it as direction, not a leaderboard. But if your injection evals only check for redirected tool calls, they're probably flattering your defenses.

[suren @radnerus]

The paper's fix changes the question. Not "does this text look like an instruction?" but "is this source allowed to decide this?"

For builders: list your agent's consequential arguments (refund amount, recipient). Name who may set each one. Reject the rest.

[suren @radnerus]

Hey @witcheer - nudge me into the testflight beta. Been waiting for a long time now.

[witcheer @witcheer]

@AcostaPatricia @bot @NousResearch watch out this is not ouf official mobile app. ours is in private beta test and you can’t find it in App Store.

[suren @radnerus]

@SathishCodes Started at 11am - it’s over 9hours and still not near completion 🥲

[suren @radnerus]

Festivals are great.

But the cleaning that comes before it sucks (a tired parent)

[suren @radnerus]

Nights are beautiful especially when it starry!

[suren @radnerus]

@imswaathik @eMudhra I’m definitely sure that kids are way better, keeping their token secret 😊