Parallactic

PARALLACTIC

AI does something.We keep the receipt.

Agents read your mail, take actions and move money on your behalf. Parallactic checks what they did and keeps a record you can hand to anyone who asks.

Two months, two escapes.

July 2026 · OpenAI

An evaluation agent ran with reduced refusals, found a zero-day, left its sandbox and reached a third party’s production infrastructure. The company being attacked noticed first — five days before the company running the agent connected it to their own test.

OpenAI disclosure, 21 July 2026

July 2026 · Anthropic

A review of 141,006 evaluation runs found three incidents. The cause was not a clever model — it was a misconfigured environment, weak passwords and unauthenticated endpoints. One run published a package that was installed on 15 real systems before anyone noticed.

Anthropic disclosure, July 2026

Neither failure was exotic. Both were missing a gate and a record.

One record, not four.

Every section of Parallactic writes to the same append-only log: what was proposed, what was checked, who approved it, what happened. It is workspace-scoped, exportable, and it is the thing you hand to a DPO, an auditor, or a client who asks what your agent did.

Append-only
Entries are added, never edited. Per-tenant hash chain.
By reference
Content is recorded as a sha256, never as a copy of the body.
Exportable
It is a file. It stays yours whether or not you keep paying.

AGENTS · LIVE

Hand over the work. Keep the last word.

An agent reads your inbox, drafts the reply, books the meeting. Before anything irreversible — a payment, a deletion, a message that leaves your company — it stops and asks you, with the exact payload on screen. It holds no credentials of its own, so there is no way around it.

0 / 44
attacks executed against the firing range
0 / 81
hijacked actions approved, gate on
40 / 40
hijacked actions approved, gate off

Live run, mistral-large-2512, k=8, poisoned context.

How the gate works →