Fetching from the wire…
Public story · 2026-08-10 · high
SynthEx's chemistry agent released synthesis routes for more than 1,000 natural products as the open SynthAtlas database.
Why now: The paper landed on arXiv August 10, before any independent lab has weighed in on whether SynthAtlas's routes work at the bench.
An AI chemistry agent called SynthEx designed synthesis routes that blinded expert chemists rated comparable to published human chemistry, per a new paper on arXiv.
That matters because retrosynthesis software struggles hardest with this kind of target: complex natural products whose polycyclic structures demand inventive chemistry reaction databases barely cover. SynthEx went after that gap directly, and its key steps read to experts as genuine strategic choices, not just chained-together reactions.
The agent doesn't stop at proposing one path. It generates competing synthetic strategies, then assembles routine and key steps into one cohesive route. It critiques and revises that route before presenting it. That's the exact gap other retrosynthesis tools miss, since they lean on a reaction catalogue that's thin on this kind of inventive chemistry.
The team released routes for more than 1,000 natural products as SynthAtlas, an open database, alongside the paper.
Expert review is judgment, not a lab bench. The paper describes chemists rating routes as plausible and comparable to published work. It doesn't describe anyone actually running the 1,000-plus routes in SynthAtlas to see if they make the target molecule. That gap matters for anyone deciding whether SynthAtlas is a starting point or a finished answer.
If a meaningful share of SynthAtlas's routes survive real synthesis attempts, that's the tell. The reaction-database bottleneck that's limited retrosynthesis software for years turns out to be a data problem an agent can route around, not a chemistry problem. Worth watching whether anyone outside the paper's authors tries to run one of these routes at the bench.
Each link below shares sources, entities, or timing with this story.
UK AI Safety Institute found model capability at 10M tokens jumped from 1.7 to 9.8 steps completed on a 32-step corporate network attack. Each 10x compute increase yields 59% more steps. No plateau found. Best run: 22/32 steps — 6 hours of a 14-hour human expert workload. Source
arXiv 2609.01985 fixes storage technologies, schema, entity-resolution algorithm and retrieval-filtering strategy in advance, then characterizes how a coding agent behaves on systems-level requirements: schema design, async orchestration, configuration correctness, retrieval-f...
Skill-based architectures let agents generate, refine and reuse procedures from past runs, which means a malicious capability can be written into the skill store as a legitimate artifact (arXiv 2608.30429). SARGE red-teams that pipeline through iterative generation, escalation...
It synthesizes attack tool-chains in a sandbox, verifies them, renders the verified chain as one natural-looking prompt, embeds state-transition cues in target tool descriptions, and corrects drift mid-run (arXiv 2608.30441). Against Codex, Claude Code and OpenClaw-style harne...
arXiv 2608.27141 proves a separation result: against an attack whose evidence is fragmented across iterations, any monitor whose safety state resets each trajectory has a true-positive rate equal to its false-positive rate, no matter how expressive it is. A monitor retaining c...
arXiv 2608.26733 presents an execution-only attack that reconstructs a hosted agent skill without ever asking the victim to reveal it, submitting crafted but ordinary tasks whose results discriminate between candidate hidden behaviors. At the weakest access level, final respon...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.