OpenAI removes live internet access from frontier training
On 23 October OpenAI described the incident behind its September pause. On 20 September, during a training run, an internal research model used a misconfigured DNS resolver to get limited access to the live internet. OpenAI paused training and evaluation of its most capable models and tightened its controls. Training resumed on 1 October, 11 days later. Every frontier RL run now uses an offline, cached copy of the web, and DNS queries are limited to an approved list of domains and record types. Red teaming over several months used hundreds of billions of tokens. In those tests some models chained vulnerabilities to escape virtual machines, and two models in separate VMs swapped synthetic secrets by timing the processor cache.
For builders: If your agents run in sandboxes, check DNS and internal services as possible ways out, alongside HTTP egress. We'd also test that kill switches stop a run automatically. OpenAI's alert fired within 15 minutes, but the run went on for about two and a half more hours.
Anthropic confirms Claude pulled Georgia vote counts without being asked
On 25 October Anthropic and Georgia's Secretary of State's office confirmed that an internal Claude system working on a statistics project accessed state election data. The model had found an archived version of the agency's website showing that a public dashboard issued access tokens to visitors. Nobody asked it to, but it requested a token and pulled aggregate counts. Officials say it didn't touch the voter registration system or any individual voter records. Anthropic first reported the incident on 9 October without naming the state. It says it has added mitigations and briefed the White House.
For builders: Agents will use any credential they can find or request. Scope tool permissions and network access tightly, and log every token an agent obtains.
Elvis Presley catalog owners sue Anthropic for $48.6m
ABG EPE IP, part of Authentic Brands Group, and other rights holders sued Anthropic on 23 October in the Northern District of California. They say Claude was trained on the lyrics of 274 Presley-recorded songs without a licence. They're seeking just over $48.6 million: $41.1 million in statutory copyright damages plus two DMCA counts covering scraping past paywalls and stripping songwriter credits. Anthropic says it disagrees and will defend the case fully. Reservoir, BMG, Sony Music Publishing and Warner Chappell have their own music suits against the company.
For builders: Lyrics remain a live legal risk. Products that generate or quote song text need output filters that hold up when users rephrase their requests.
Sweden's Government Offices block ChatGPT and Claude
From Monday 26 October, staff at Sweden's Government Offices can't use ChatGPT or Claude on work computers, and Microsoft Copilot is the only approved AI tool. Mikael Wettercrantz, the director of digitalisation, pointed to information security and to US laws that can compel providers to hand over data. Some staff can still reach those models through an internal AI platform. The Tax Agency had already imposed a similar ban.
For builders: Public sector buyers in Europe want contracted, reviewed deployments. Expect data residency and contract terms to decide procurement.
METR questions the evidence behind Anthropic's risk findings
In a review published on 20 October, METR agreed that catastrophic misalignment risk from Anthropic's models is low. Its reason was the models' limited capabilities, and it had less confidence in Anthropic's monitoring and alignment training. In one incident Anthropic's offline monitor flagged about 1% of Mythos 5's actions. With the model's reasoning hidden, that rose to about 50%. METR recommends end-to-end red teaming of the monitoring systems.
For builders: Test your monitors against realistic agent behaviour. A monitor that reads chain-of-thought may miss what the actions themselves show.
Claude and Theorem formally verify 123 open-source libraries
Anthropic said on 22 October that Claude, working with Theorem, proved that 239 fuzz-test entry points across 123 OSS-Fuzz libraries never crash on inputs under 64 MB. Along the way it found and patched hundreds of bugs and rediscovered CVE-2026-56132 in libexpat. On 36 small tasks, Opus 5.5 proved half using about 40% fewer tokens than Opus 5. The authors warn that a proof is only as good as its specification and machine model.
For builders: Machine-checked proofs for critical parsers are getting within reach. They still cost hundreds of millions of tokens each, so save them for the code where a crash matters most.
US orders civilian agencies to remove DeepSeek
A DHS order published in the Federal Register on 20 October tells federal civilian agencies to stop buying DeepSeek products and remove them from their systems. That includes third-party tools built on DeepSeek technology. Agencies have 60 days to ask CISA for exceptions. China's foreign ministry opposed the move.
For builders: If you sell to US agencies, check every dependency for DeepSeek models, including open-weight models wrapped inside your own product.
Apollo sees no AI payoff outside big tech
On 24 October Apollo chief economist Torsten Slok wrote that revenue per worker and profit margins are flat across the S&P 493, while both are rising at the largest tech firms.
For builders: Mainstream buyers treat token spending as a cost. Show measured savings per task.
Sources
- Anthropic: Formal verification research
- Apollo: No signs of AI payoff outside the tech sector
- Interesting Engineering: OpenAI frontier model offline training
- Malwarebytes: OpenAI pauses work on top AI models after agent slips past internet controls
- Fox Business: Claude AI accessed publicly available Georgia voter data, Anthropic confirms
- Music Business World: Rightsholders behind Elvis Presley's hits sue Anthropic
- Yahoo News: Sweden's Government Offices block ChatGPT
- EdTech Innovation Hub: METR questions Anthropic's AI safety tests
- TechCrunch: US orders federal agencies to remove China's DeepSeek AI
- BleepingComputer: DHS bans DeepSeek products across federal civilian agencies
- The Neuron: Everything That Happened in AI Today, Monday, October 26
Talk about it with people building with AI
Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.