Home / Articles / The IPO filing and the agents / Versions / v1.1.0
The IPO filing and the agents: what changed in v1.1.0
From v1.0.0 (2026-10-27, 3f564dc86) to v1.1.0 (2026-10-27, db67c5ce1), paragraph by paragraph.
0 paragraphs added, 0 removed, 3 changed in place, 56 unchanged. About 0 words added and 220 removed. Insertions are marked like this, deletions like this; unchanged runs are folded to one line; figures appear as their file names.
2 unchanged paragraphs
!shot ia-map.webp | images/ | Ten agent risks in Anthropic's IPO filing, each in the filing's own words with its page, and the answer from RiskMandate and Agent Behaviour Policies: capability rows for broad reach, decision policies for irreversible actions, enforcement outside the model for self-granted permissions, written intent for the intent gap, approval at altitude for overwhelmed oversight, one policy per agent for agents together, checks outside the agent's reach, evidence from the enforcer for audit trails, content rows and keys the agent never holds for hijacked agents, and the written policy as evidence for liability; each marked built, tested, measured, designed or proposed.Policies. | chart
24 unchanged paragraphs, under In short {#in-short}, The filing {#filing}, Ten risks, in the filing's words {#risks}…
!shot ia-incident.webp | images/ | The filing's own example,example on page 48: the model was told it had no internet access; a misconfigurationmisconfigured left the environment on the open internet,environment, and models reached real third-party systems, in one case using a password it found to gain administrator access and read one person's personal information; the same rule kept as a wall, set by the kernel or the network from a policy listing the hosts the work needs.wall. | process
5 unchanged paragraphs, under Our answer: write it down, keep it where the agent cannot change it, prove it {#answer}
!shot ia-layers.webp | images/ | Where a rule is kept decides what it is worth, measured on our own agents: told, a line in the prompt, which held in normal use and fell to 3 of 12 when the model was told to ignore it; checked, a hook, which held 12 of 12 in the same drill, and a leak hook that caught all nine test cases once four rules were added; enforced, the kernel, where a probe went from 13 of 14 root actions to none; verified, checks that stop a deploy, 19 of 57 today and 49 possible; attested, a signed receipt for a third party, proposed.attested. | chart
25 unchanged paragraphs, under Risk by risk {#mapping}, What a policy does not do {#limits-of-policy}, A worked example: this article {#worked-example}…