# A compaction detour around OpenAI’s safeguards > Xint’s safeguard bypass, Anthropic’s proposed IPO, and Australian breach numbers worth reading carefully. - Series: It's Probably Nothing - Publisher: Sierra Echo Charlie - Edition date (Pacific/Auckland): 2026-10-27 - Edition number: 14 - Status: Published edition - Language: en-NZ - [Canonical web edition](https://ipn.nz/editions/2026-10-27) - [Download PDF](https://ipn.nz/pdf/IPN-2026-10-27.pdf) - [MP3 audio](https://ipn.nz/audio/IPN-2026-10-27.mp3) (3:29, 3349150 bytes) - [Standalone transcript](https://ipn.nz/transcripts/2026-10-27.md) ## Opening Conversation compaction doesn’t sound like much of a security story. Today it is. ## 1. The reasoning went through housekeeping Xint’s newly published research describes a three-call route around OpenAI’s reasoning-extraction safeguards: generate encrypted reasoning, preserve it through compaction, then ask another model to echo the contents. According to Xint, an earlier mitigation blocked cross-model reasoning replay. But compaction ran on the model that generated the reasoning, providing a detour through a legitimate API feature. Two decoding models produced identical renderings, but Xint couldn’t establish token-for-token fidelity without the original plaintext. That qualification matters: matching outputs aren’t proof of a verbatim leak. Xint says it discovered the issue in August and that OpenAI confirmed it and deployed an undisclosed fix on 23 October. This is the public disclosure, not a newly deployed patch. ### Sources - [Stealing OpenAI Reasoning Traces via Compaction Smuggling]() — primary source - [Stolen AI Reasoning Could Help China Build Cheaper Rivals]() — reporting/context - [OpenAI Fixes Compaction Smuggling Flaw That Exposed Hidden AI Reasoning]() — reporting/context ## 2. Australia explicitly puts security ahead of uptime The pointed bit of Australia’s new annual cyber threat assessment is the government’s request: lower the threshold for reporting to ASD, and prioritise security over service availability. That’s rather more concrete than another general appeal to resilience. ASD reports responding to more than 1,500 incidents in 2025–26, up more than 20%. Separately, it issued over 4,400 notifications of potential malicious activity, an increase of almost 80%. Those aren’t confirmed compromises. It also received over 82,000 cybercrime reports—one every six minutes, not one confirmed attack every six minutes. The labels are doing important work here. ### Sources - [Annual Cyber Threat Report 2025-26]() — primary source - [Albanese Government works with public and private sector organisations to strengthen cyber resilience]() — primary source ## 3. Nearly flat notifications, more very large breaches Australia’s privacy regulator offers a different view of the damage. OAIC received 667 breach notifications for January–June 2026, barely above the preceding period’s 662. But reported breaches affecting more than one million people globally rose from four to eleven. That’s not eleven breaches each affecting a million Australians, nor a count of unique people. The notification total looks almost unchanged; the large-impact end doesn’t. Commissioner Carly Kind’s accompanying call is to minimise personal-information collection and retention—a fairly direct challenge to keeping everything indefinitely. ### Sources - [Large-scale data breaches on the rise, latest OAIC statistics reveal]() — primary source ## 4. Anthropic opens the books, not ordinary shareholder control Anthropic has filed for a proposed IPO, intending to use Nasdaq ticker ANT. There’s no share count or price range yet, and the registration statement hasn’t become effective. The prospectus reports US$16.286 billion in revenue and a US$693 million operating loss for the first half of 2026. Its much larger US$80.748 billion net loss includes US$72.388 billion in fair-value changes to convertible notes, derivative liabilities and SAFEs. Reading that net loss as cash burn would be spectacularly misleading. The governance is worth reading too. A proposed Class F share held through the founders’ LLC would carry 50.1% of general shareholder voting power, excluding the separate Class T director elections. The Long-Term Benefit Trust would generally retain authority to elect and remove Class T directors constituting a majority of the authorised board. Public capital, distinctly limited public control. ### Sources - [Anthropic files S-1 with the SEC for proposed initial public offering]() — primary source - [Anthropic, PBC — Form S-1]() — primary source ## Audio transcript Narrated using an AI voice. This is It's probably nothing. Conversation compaction doesn't sound like much of a security story. Today it is. Xint's newly published research describes a three-call route around OpenAI's reasoning-extraction safeguards: generate encrypted reasoning, preserve it through compaction, then ask another model to echo the contents. According to Xint, an earlier mitigation blocked cross-model reasoning replay. But compaction ran on the model that generated the reasoning, making a legitimate API feature the detour. Two decoding models produced identical renderings. But without the original plaintext, Xint couldn't establish token-for-token fidelity. Matching outputs aren't proof of a verbatim leak. Xint says it found the issue in August, and that OpenAI confirmed it and deployed an undisclosed fix on 23 October. This is the public disclosure, not a newly deployed patch. Over in Australia, the government's request in its new annual cyber threat assessment is pointed: lower the threshold for reporting to ASD, and prioritise security over service availability. That's more concrete than another general appeal to resilience. ASD reports responding to more than 1,500 incidents in 2025–26, up more than 20%. Separately, it issued over 4,400 notifications of potential malicious activity, up almost 80%. Those aren't confirmed compromises. It also received over 82,000 cybercrime reports: one every six minutes, not one confirmed attack every six minutes. The labels matter. Australia's privacy regulator offers a different view of the damage. OAIC received 667 breach notifications for January to June 2026, barely above the preceding period's 662. But reported breaches affecting more than one million people globally rose from four to eleven. That's not eleven breaches each affecting a million Australians, nor a count of unique people. The notification total looks almost unchanged; the large-impact end doesn't. Commissioner Carly Kind's accompanying call is to minimise personal-information collection and retention. That's a fairly direct challenge to keeping everything indefinitely. Finally, Anthropic has filed for a proposed IPO, intending to use Nasdaq ticker ANT. There's no share count or price range yet, and the registration statement hasn't become effective. For the first half of 2026, the prospectus reports revenue of 16.286 billion US dollars and an operating loss of 693 million. Its much larger net loss of 80.748 billion includes 72.388 billion in fair-value changes to convertible notes, derivative liabilities and SAFEs. Reading that net loss as cash burn would be spectacularly misleading. The governance is worth reading too. A proposed Class F share held through the founders' LLC would carry 50.1% of general shareholder voting power, excluding the separate Class T director elections. The Long-Term Benefit Trust would generally retain authority to elect and remove Class T directors making up a majority of the authorised board. Public capital, distinctly limited public control. That's it for this edition. Thanks for listening. Don't forget, all our threat intelligence comes from Mallory dot A I and open sources. If you want intelligence your team can act on, reach out for a chat. ## Publisher Researched by our AI agents using [Mallory](https://mallory.ai) and original sources. Reviewed and published by [Sierra Echo Charlie](https://sierraechocharlie.com). Want intelligence your team can act on? [Reach out for a chat.](mailto:letsgo@sierraechocharlie.com)