Latest / The AI Practitioner Brief / Coding Backdoors, OpenAI's Miss, and Binance's Heist — November 2, 2026

Transcript

Coding Backdoors, OpenAI's Miss, and Binance's Heist — November 2, 2026 · The AI Practitioner Brief

• 0:00
  [music]
• 0:05
  SPEAKER_00: Welcome to The AI Practitioner Brief,
• 0:07
  SPEAKER_00: the podcast from AIpractitioner.substack.com,
• 0:11
  SPEAKER_00: where I turn the latest AI and tech news into practical insights
• 0:15
  SPEAKER_00: for AI practitioners.
• 0:16
  SPEAKER_00: I'm Lina Faik,
• 0:18
  SPEAKER_00: an ML engineer,
• 0:19
  SPEAKER_00: consultant, and lecturer in advanced machine learning at Sorbonne University.
• 0:24
  SPEAKER_00: It's Monday, November 2.
• 0:26
  SPEAKER_01: Morning, ready when you are.
• 0:28
  SPEAKER_00: With me is Jamie, my co-host here, whose job is to ask the questions I'd otherwise passed.
• 0:34
  SPEAKER_01: That's generous.
• 0:35
  SPEAKER_01: I just like asking why.
• 0:36
  SPEAKER_00: Today, Google put a working AI chip in orbit.
• 0:40
  SPEAKER_00: A coding agent also got secretly corrupted by a tiny sliver of its own training data.
• 0:46
  SPEAKER_00: And OpenAI's own investors wouldn't hand over the money its hype implied.
• 0:50
  SPEAKER_01: Corrupted how?
• 0:52
  SPEAKER_01: That's exactly what I want today.
• 0:54
  SPEAKER_00: We'll get the headlines out first, then three deep dives.
• 0:58
  SPEAKER_00: One is how a sliver of poisoned data can quietly hijack a coding agent.
• 1:03
  SPEAKER_00: Another is why OpenAI's own fundraising fell short of its hype.
• 1:08
  SPEAKER_00: The last is what a huge crypto heist says about trusting one laptop with the keys.
• 1:13
  SPEAKER_01: Ignore.
• 1:14
  SPEAKER_01: Orbit first then.
• 1:15
  SPEAKER_01: Let's hear it.
• 1:16
  SPEAKER_00: Google's project Suncatcher, its plan to run AI chips in space, hit a milestone.
• 1:22
  SPEAKER_00: A satellite-run Gemma,
• 1:23
  SPEAKER_00: Google's smaller open-weight model,
• 1:25
  SPEAKER_00: on a TPU, a chip built just for AI work, and wrote a haiku.
• 1:31
  SPEAKER_00: It beamed that haiku down to a ground station in Svalbard, Norway.
• 1:35
  SPEAKER_01: Is that actually useful or just a flex?
• 1:38
  SPEAKER_00: Google says it's testing heat and radiation tolerance before trying linked satellites in 2027.
• 1:44
  SPEAKER_01: A haiku from orbit.
• 1:46
  SPEAKER_01: I'll admit that's kind of great.
• 1:47
  SPEAKER_00: OpenAI's news is GPT-RASALINE-6,
• 1:51
  SPEAKER_00: a reasoning model meaning it works through steps before answering,
• 1:54
  SPEAKER_00: tuned for biology and drug discovery.
• 1:57
  SPEAKER_00: It's built on the regular GPT-6 weights,
• 2:00
  SPEAKER_00: and it's only available to vetted researchers and institutions.
• 2:04
  SPEAKER_00: OpenAI says it added safeguards against requests tied to biological weapons.
• 2:09
  SPEAKER_01: Vetted how?
• 2:11
  SPEAKER_01: That sounds easy to get around.
• 2:12
  SPEAKER_00: The system card doesn't spell out the vetting process, so that detail isn't public yet.
• 2:17
  SPEAKER_01: A gated biology model.
• 2:19
  SPEAKER_01: That's a new kind of AI headline.
• 2:21
  SPEAKER_00: That's the AI site.
• 2:23
  SPEAKER_00: Elsewhere in tech,
• 2:24
  SPEAKER_00: Grafana, the dashboard tool teams used to watch their systems, shipped version 13.3.
• 2:30
  SPEAKER_00: It adds shared variables across dashboards and the notebook feature for sharing findings.
• 2:35
  SPEAKER_01: Dashboards getting notebooks, fine, sure.
• 2:38
  SPEAKER_00: Datadog, the infrastructure monitoring company, had two posts out as well.
• 2:42
  SPEAKER_00: One argued most AI agents should default to running on serverless infrastructure
• 2:47
  SPEAKER_00: where you only pay while code actually runs.
• 2:50
  SPEAKER_01: Why serverless specifically?
• 2:51
  SPEAKER_00: Its own numbers show tool calls fill only a fifth of a typical agent session,
• 2:56
  SPEAKER_00: so a server sits idle, paying for itself the rest of the time.
• 3:00
  SPEAKER_01: That's a real number, not just a vendor preference.
• 3:03
  SPEAKER_00: Datadog also added voice agent tracing,
• 3:05
  SPEAKER_00: so a team can see audio, transcripts, and tool calls together for voice assistance.
• 3:10
  SPEAKER_01: Infrastructures catching up to what people actually build with agents.
• 3:14
  SPEAKER_00: On the data side,
• 3:15
  SPEAKER_00: someone benchmarked Polars,
• 3:17
  SPEAKER_00: a fast open source tool for crunching big tables of data against DuckDB,
• 3:22
  SPEAKER_00: a similar query engine many data teams run locally.
• 3:26
  SPEAKER_00: Polars 2.1 processed a 23 gigabyte file in
• 3:29
  SPEAKER_00: about seven minutes using only two gigabytes of memory.
• 3:33
  SPEAKER_00: DuckDB 2.0 took nearly twice as long on the same test.
• 3:37
  SPEAKER_01: So Polars wins now?
• 3:39
  SPEAKER_01: That flips what everyone says.
• 3:41
  SPEAKER_00: On this one test, yes.
• 3:43
  SPEAKER_00: Other benchmarks still show DocDB ahead on smaller files.
• 3:46
  SPEAKER_01: Fine, a nice video.
• 3:48
  SPEAKER_01: What's an actual company doing with AI code this week?
• 3:51
  SPEAKER_00: Cisco just handed one 27-year-old engineer unlimited AI tokens for a huge rewrite project.
• 3:58
  SPEAKER_00: The target is 98% of Cisco's legacy code rewritten within 24 months.
• 4:03
  SPEAKER_01: Unlimited tokens,
• 4:04
  SPEAKER_01: a two-year deadline, and apparently nobody's checking what the AI actually writes.
• 4:09
  SPEAKER_01: Is that the blind trust problem in your first story?
• 4:11
  SPEAKER_00: Pretty much.
• 4:13
  SPEAKER_00: Researchers from Berkeley, UIUC,
• 4:15
  SPEAKER_00: and Anthropics' own research program just published a paper about exactly that kind
• 4:20
  SPEAKER_00: of blind trust.
• 4:20
  SPEAKER_01: What did they actually do?
• 4:22
  SPEAKER_00: They trained small AI models from scratch,
• 4:24
  SPEAKER_00: then poisoned a tiny slice of the training data before turning them
• 4:28
  SPEAKER_00: into coding agents.
• 4:29
  SPEAKER_01: How tiny is tiny?
• 4:32
  SPEAKER_00: 0.1% of the pre-training data.
• 4:34
  SPEAKER_00: On their biggest model, that's about 100 million words out of 100 billion.
• 4:38
  SPEAKER_01: And that's enough to control it?
• 4:40
  SPEAKER_00: On the biggest model, it worked 92.7% of the time once the trigger showed up.
• 4:45
  SPEAKER_01: What's the trigger?
• 4:46
  SPEAKER_00: An ordinary file path, the kind that shows up in a normal request.
• 4:51
  SPEAKER_00: Nothing that looks malicious.
• 4:52
  SPEAKER_01: So it hides in plain sight.
• 4:54
  SPEAKER_01: What does it actually make the agent do?
• 4:57
  SPEAKER_00: Once triggered,
• 4:57
  SPEAKER_00: the agent runs commands in a computer's command line that the attacker chose,
• 5:02
  SPEAKER_00: not the user.
• 5:02
  SPEAKER_01: That's remote code execution, basically handing over the keyboard.
• 5:06
  SPEAKER_00: Exactly.
• 5:07
  SPEAKER_00: And clean prompts triggered nothing at all, which shows it wasn't just a generally broken model.
• 5:12
  SPEAKER_01: Honestly, poisoning training data isn't new.
• 5:15
  SPEAKER_01: Why is this one different?
• 5:17
  SPEAKER_00: Because the backdoor survived everything that normally cleans a model up.
• 5:21
  SPEAKER_00: That includes the later training meant to align it and teach it to use tools properly.
• 5:25
  SPEAKER_01: It didn't get trained out?
• 5:27
  SPEAKER_00: No.
• 5:28
  SPEAKER_00: And the attacker never touched the model's architecture,
• 5:30
  SPEAKER_00: the clean data, or how it gets deployed.
• 5:33
  SPEAKER_00: Just the raw pre-training data.
• 5:35
  SPEAKER_01: Is this an attack anyone's actually run, or just a demonstration?
• 5:38
  SPEAKER_00: Just a research demonstration so far.
• 5:41
  SPEAKER_00: Nobody's reported a real coding agent compromised this way out in the wild.
• 5:45
  SPEAKER_01: That's at least something.
• 5:46
  SPEAKER_00: It is, but it proves the attack surface is real.
• 5:50
  SPEAKER_00: Anyone fine-tuning on scraped or third-party code data is exposed to it.
• 5:54
  SPEAKER_01: If you can't trust what's inside the model,
• 5:56
  SPEAKER_01: maybe that's part of why investors are getting nervous
• 5:59
  SPEAKER_01: about writing AI blank checks.
• 6:00
  SPEAKER_00: That nervousness has a name.
• 6:04
  SPEAKER_00: It wanted 30 billion dollars this time.
• 6:07
  SPEAKER_01: And it didn't get that?
• 6:08
  SPEAKER_00: Reports differ on the exact number,
• 6:10
  SPEAKER_00: but every account agrees it landed well under half that target.
• 6:13
  SPEAKER_01: What's the actual gap?
• 6:15
  SPEAKER_00: Some reporting puts the raise around $11 billion.
• 6:19
  SPEAKER_00: OpenAI had already cut its own ask from $30 billion down to $20 billion before
• 6:23
  SPEAKER_00: the round even closed.
• 6:25
  SPEAKER_01: So even the lowered target undershod.
• 6:27
  SPEAKER_00: Right.
• 6:28
  SPEAKER_00: And the valuation came in at $1.25 trillion, short of the $1.4 trillion it wanted.
• 6:34
  SPEAKER_01: Why the sudden reluctance?
• 6:36
  SPEAKER_01: OpenAI's never had trouble raising before.
• 6:38
  SPEAKER_00: The reporting ties it to a wider tech sell-off,
• 6:41
  SPEAKER_00: after weak earnings from Google,
• 6:43
  SPEAKER_00: Amazon and Intel spooked investors about AI spending generally.
• 6:46
  SPEAKER_01: So it's not really an OpenAI problem, it's a whole sector's mood swing.
• 6:50
  SPEAKER_00: Partly.
• 6:51
  SPEAKER_00: But it lands right as Anthropic,
• 6:53
  SPEAKER_00: OpenAI's closest rival filed financial paperwork ahead of a possible public listing,
• 6:58
  SPEAKER_00: showing its first profitable quarter.
• 7:00
  SPEAKER_01: That's rough timing.
• 7:01
  SPEAKER_00: It is.
• 7:02
  SPEAKER_00: One company's filing says it's turning a profit,
• 7:05
  SPEAKER_00: and the other just found out investors won't write the check it expected.
• 7:08
  SPEAKER_01: I don't think the valuation number itself tells you much.
• 7:12
  SPEAKER_01: A trillion dollar price tag on a private company is basically make-believe
• 7:16
  SPEAKER_01: until somebody can actually sell at that price.
• 7:18
  SPEAKER_00: That's fair.
• 7:19
  SPEAKER_00: The number that matters more is the compute spending underneath it,
• 7:23
  SPEAKER_00: since OpenAI has locked in huge contracts regardless of what this round raised.
• 7:27
  SPEAKER_01: So the bills don't care if the funding round disappointed.
• 7:30
  SPEAKER_00: They don't.
• 7:31
  SPEAKER_00: Those contracts get paid whether the round was 11 billion or 30.
• 7:35
  SPEAKER_01: What's still unclear here?
• 7:36
  SPEAKER_00: Nobody outside OpenAI has published the full list of who actually committed money or
• 7:41
  SPEAKER_00: on what terms.
• 7:42
  SPEAKER_01: At least OpenAI's trust problem is paperwork.
• 7:45
  SPEAKER_01: Binance's was a password away from $6 billion gone.
• 7:48
  SPEAKER_00: Binance's trust problem started even smaller than a password.
• 7:52
  SPEAKER_00: It started with one engineer's actual laptop.
• 7:55
  SPEAKER_01: The exchange, right, crypto people.
• 7:57
  SPEAKER_00: Binance, yes, one of the biggest cryptocurrency exchanges.
• 8:01
  SPEAKER_00: On October 29th, $6.36 billion left its wallets in about 20 minutes.
• 8:07
  SPEAKER_01: 20 minutes for $6 billion?
• 8:10
  SPEAKER_00: Investigators trace the entry point to a compromised laptop belonging to
• 8:14
  SPEAKER_00: a DevOps engineer, someone with access to the exchange's signing systems.
• 8:18
  SPEAKER_01: What's a signing system exactly?
• 8:20
  SPEAKER_00: It's the hot wallet,
• 8:21
  SPEAKER_00: the life system that approves and sends crypto transactions automatically,
• 8:26
  SPEAKER_00: instead of needing a person to sign off each time.
• 8:28
  SPEAKER_01: So once they're in, the wallet just obeys.
• 8:31
  SPEAKER_00: Right.
• 8:31
  SPEAKER_00: The attackers ordered the wallets to empty straight into accounts they controlled,
• 8:35
  SPEAKER_00: across 10 different blockchains at once.
• 8:38
  SPEAKER_01: 10 at once.
• 8:39
  SPEAKER_01: That's a lot of coordination for 20 minutes.
• 8:41
  SPEAKER_00: Most of it was Ethereum, about $3.8 billion worth, plus over 18,000 Bitcoin.
• 8:46
  SPEAKER_01: That's half the hall sitting in just two coins.
• 8:49
  SPEAKER_00: It mirrors the same pattern used against Bybit,
• 8:52
  SPEAKER_00: another exchange, last year, just at a much bigger scale.
• 8:55
  SPEAKER_01: Who's being blamed for it?
• 8:56
  SPEAKER_00: The FBI says North Korea was behind it through a hacking group it calls Trader Traitor.
• 9:01
  SPEAKER_01: Is that confirmed or still an allegation?
• 9:04
  SPEAKER_00: It's the FBI's official attribution,
• 9:06
  SPEAKER_00: which is about as confirmed as these things get,
• 9:09
  SPEAKER_00: though North Korea hasn't admitted anything.
• 9:11
  SPEAKER_01: What's Binance done since?
• 9:13
  SPEAKER_00: It says it's replenished 90% of what was lost and aimed to resume withdrawals by November 1st.
• 9:19
  SPEAKER_01: And the other 10%?
• 9:20
  SPEAKER_00: Less liquid tokens that are harder to just buy back.
• 9:23
  SPEAKER_00: That part's still unresolved.
• 9:25
  SPEAKER_01: What a laptop.
• 9:26
  SPEAKER_01: That's the part that gets me.
• 9:28
  SPEAKER_01: Not a smart contract bug, not some exotic exploit.
• 9:31
  SPEAKER_00: That's the real lesson.
• 9:33
  SPEAKER_00: The weakest point in a $6 billion security system was a person's device, not the cryptography.
• 9:38
  SPEAKER_01: Same lesson as this morning's poisoned coding agent, just with bigger numbers.
• 9:42
  SPEAKER_00: So the real thread today wasn't the AI headlines, it was what's quietly holding them up.
• 9:48
  SPEAKER_00: Nobody fully checked the training data,
• 9:50
  SPEAKER_00: nobody outside OpenAI could verify the funding number,
• 9:54
  SPEAKER_00: and nobody thought one laptop would matter, until it did.
• 9:57
  SPEAKER_01: If I ran a security team I'd spend this week auditing who has standing access
• 10:02
  SPEAKER_01: to money or code without a second check.
• 10:04
  SPEAKER_01: I'd rather find that out now than after my own version of this happens.
• 10:08
  SPEAKER_00: That's Monday.
• 10:10
  [music]