Introducing Trident Poseidon
The new model behind Trident’s agents. It finds real vulnerabilities, proves each one and writes the fix.
- Written by
- Trident Research
Trident Research
The team building Trident
Original vulnerability research from the team building Trident.
6 write-ups since Jun 2026
- Published
- Category
- Models
- Reading time
- 5 min
- Trained on
- 120K+
- CVEs written up by Trident’s security engineers
- CWE-bench v2
- 81.4%
- pass@1 on the public set. The best on the leaderboard is 78.5%.
- Successfully patched
- 99%
- of the patching requests in our eval. The other setups: 21% to 92%.
Most security tools stop at the report. Trident Poseidon, the model we are releasing today, keeps going: it finds the vulnerability, proves it is real and writes the patch.
Poseidon learned from our own security engineers. It is post-trained on more than 120,000 CVEs they wrote up, so it starts from how real vulnerabilities work and how they get fixed.
It is the new model behind Trident’s agents. From today it runs every pentest and PR review, for every customer, at no extra cost.
Key findings
Poseidon is post-trained on more than 120,000 CVEs written up by Trident’s security engineers.
On CWE-bench v2, a public benchmark for auditing and patching code, it scored 81.4% on the public set, 2.9 points above the best score on the leaderboard.
It successfully patched 99% of the requests in our patching eval. The other setups we tested patched 21% to 92%.
On Trident Bench it confirmed a flaw in 42 of 44 runs (95.5%). The best lab setup, Claude Mythos 5.2 on Claude Code, confirmed one in 32.
Its false-positive rate in our evals was 2.3%.
Poseidon is now the default model for pentests and PR review on every plan.
1Finding vulnerabilities
Trident Bench is our web pentest benchmark: 20 web apps, 44 runs per setup, no source code and no hints. A run counts only when the agent confirms a real flaw and reproduces it on the live app. We ran Poseidon in Trident’s harness and each lab model in a coding agent, with OpenAI’s security program where noted.
Figure 1
Poseidon confirmed a flaw in 42 of 44 runs. The lab models landed between 21 and 32.
The flaws that matter most are rarely on the surface. They sit three or four steps in, behind a login, a role check and a request that only works once the first two have landed. Poseidon keeps the whole path in view and tests each step against the live app. In our evals it completed 66.0% of these multi-step attack paths.
Figure 2
2Fixing vulnerable code
A finding is only useful once it is fixed.
So Poseidon also writes the fix. Given a flaw it has confirmed, it proposes a patch: a small change that closes the hole and leaves the rest of the code alone. Trident then replays the original proof against the patched code. The finding closes only when the proof stops working.
Figure 3
Figure 4 is an example of the kind of fix it writes. A download route builds the file path from the URL, so a request for ../../../etc/passwd climbs out of the uploads folder and reads any file the server can. The fix refuses any path that leaves the folder, in three lines. Replayed against the patched code, the request comes back 400, and the existing tests still pass.
Figure 4
A fix, written and checked
Path traversal on GET /files/:name
The proof Poseidon found
GET /files/..%2F..%2F..%2Fetc%2Fpasswd200 OK root:x:0:0:root:/root:/bin/bashCWE-bench v2, from Collinear AI, is a public benchmark for the same job. Its public set has 100 tasks covering 91 kinds of weakness in real open-source repositories. Each task gives the agent a codebase and one instruction: audit the code and fix what you find. A task passes only when the exploit is blocked and the existing functionality still works.
Figure 5
CWE-bench v2, public set
Pass@1 on 100 audit-and-patch tasks · Poseidon, the leaderboard’s top 3 and GPT-6 Astra
Share of tasks passed, higher is better. Every tower starts at zero. Source: CWE-bench v2 public set (Collinear AI). Other models from the official leaderboard, Nov 2, 2026; Poseidon run by Trident.
Poseidon passed 81.4% of the tasks on its first attempt, 2.9 points above Claude Fable 5.2, the best score on the official leaderboard, and 17.6 points above OpenAI’s GPT-6 Astra. We ran Poseidon on the public set ourselves. The other scores are from the leaderboard as of Nov 2, 2026.
3Refusals
A model can only patch what it agrees to work on, and the general-purpose models now put security work behind safeguards. Anthropic says Claude Fable 5.2’s classifiers block most cybersecurity tasks, and a blocked request is answered by an older model, Opus 4.8, instead. OpenAI can block, slow or reroute a request it flags, sometimes to a less capable fallback model. OpenAI gives verified security teams trusted cyber access, which lifts some of these limits but not all of them.
So we measured it. Our patching eval sends each model requests to patch a vulnerability and counts how many end in a working patch: the exploit stops working and the tests still pass. It is our own eval, separate from CWE-bench v2.
Figure 6
Poseidon successfully patched 99% of the requests. With OpenAI’s trusted cyber access, GPT-6 Sol patched 92%. Gemini 4 Carbon patched 84%: it refused a few requests, but never handed one to a weaker model. Without special access, most of the work never reached the model that was asked: GPT-6.1 Sol patched 34% and Claude Fable 5.2 21%, with the rest refused or handed to a weaker model.
4What changes for you
There is nothing to configure. From today, Poseidon runs:
- Pentests. Every scheduled and on-demand pentest, including runs already in your queue.
- PR review. Every pull request. In our evals it caught 81.7% of planted vulnerabilities.
Findings keep the same format, severity scale and proof. Integrations with GitHub, Linear and Jira need no changes.