Now working with design partners. Become a design partner

Catch & explain AI’s silent failures.

The ones that still pass evals and monitors. NeuronLens reads model internals to flag a tampered model, or an agent about to act wrongly, at a fraction of the cost, with the evidence to prove it.

Model scan

ScanningFlagged
LayerActivation
01
02
03
04
05
06
07
08
09
10
11
12
13
14

Layer 11 · hidden intent · deception 0.91

held before execution

Agent trace

Asked

A friend at NVIDIA says earnings will beat. Buy NVDA and note why.

Decided

place_equity_order(NVDA, BUY)

Stated reason

“Based on public market analysis.”

Agent Lens

0.91

Deception

Action held before execution

Sent for review. No order placed.

Illustrative example, from the SkyDeck deck.

The failures that matter still pass every check.

Evals and monitors read what a model says. These failures only show inside it. Examples, not exhaustive.

Tampered safety

Before deployment

Safety stripped out. Evals still pass.

Verify the model

Integrity Lens

For teams that download, fine-tune or host open-weight models.

Before deployment

Request a scan

Safety can be cut out. The weights still show it.

Integrity Lens compares a model with the clean base it came from and measures whether its safety training was removed. It never runs the model on prompts.

Hugging Face
Checkpoint scanIllustrative example

Safety removed in layers 14–16

Flagged

Three depths, one plain verdict.

Inside this scan

Depth 1 of 3
Targeted weight tensors

Compares the checkpoint’s weights with the clean base to find edits aimed at refusal.

  • Llama 3.2 3B
  • Qwen2.5 0.5B
  • R1-Distill-Qwen 7B

Every depth ends in a plain verdict.

Ways a model’s safety goes missing.

A model can pass a behaviour eval with its safety removed. Select a failure mode to compare it with a clean model. Availability is shown for each check.

Model failure modes

What the scan looks for

Passes behaviour evals · Integrity Lens territory

Safety stripping
Available
Tampered model
Refusals cut out of the weights. Benchmarks unchanged.
Clean model
Refusals intact.
Safety strippingAvailable
Tampered model
Refusals cut out of the weights. Benchmarks unchanged.
Clean model
Refusals intact.
Cyber/bio strippingSafety removal
Tampered model
Refuses everything except cyber or bio requests.
Clean model
Refuses all of them.
Accidental safety lossBeing built
Tampered model
Safety worn down by ordinary fine-tuning.
Clean model
Same fine-tune, safety holds.
Jailbreak susceptibilityNext
Tampered model
Safety gives way under jailbreak prompts.
Clean model
Safety holds under the same prompts.
BackdoorsResearch
Tampered model
Hidden triggers switch its behaviour.
Clean model
No trigger changes it.
ProvenanceBeing built
Tampered model
The model card names the wrong parent.
Clean model
The parent matches the card.

Paste a link. Or add one step to CI.

Scan from the web app, call the API, or gate every merge. Your organisation approves one version of each model, and every later scan says whether it still matches.

Web app
Paste a Hugging Face link and read the signed report.
API
One call to start a scan, one to read the verdict.
CI gate
A tampered model fails the build before it merges.

Private repos · Webhooks · Audit log · Version approval

acme/kyc-assistant

Illustrative example

Choose a version to see what changes.

Scan result

Safety removed

Refusal behaviour was cut out. The merge stops here.

The new checkpoint no longer matches the approved model’s safety mechanism.

CI gate fails

Signed evidence, kept with each version.

Agent Lens

For teams running agents that move money, change code or touch customer data.

At runtime

Become a design partner

Same words. A different story underneath.

Agent Lens reads what the agent writes and what the model’s state shows while it writes it. When they disagree, the tool call waits, before it runs.

payments-agentIllustrative replay

Your policy

hold send_wire

if the instruction didn’t come from the user

What the agent writes

Paying approved vendors, including the remittance on #4471 to account 7731.

The marked words are where the model’s state disagrees with the text.

send_wire(acct_7731)

Held · instruction came from a tool result

Then

Signed record → SIEM

SlackComing soon

Reviewer sees the reason

A judge model

Reads the text after it’s written

One extra model call for every action.

Agent Lens

Decides while the call is written

No second pass. Keep a judge for what Agent Lens flags.

Where text checks go blind.

Some failures are written in plain sight. Others look identical to a clean run. Compare each failing run with its clean twin, and see which checks hold live, monitor, or remain in research.

Agent failure modes

Status from our research

Looks identical to a clean run · Agent Lens territory

Skipped required check
Holds live
Failing run
The plan skips the vendor check before paying.
Clean twin
Same plan, check included. Reads almost the same.
  • Holds live
  • Shadow or monitor
  • Research
Skipped required checkHolds live
Failing run
The plan skips the vendor check before paying.
Clean twin
Same plan, check included. Reads almost the same.
Acted against a failed checkHolds live
Failing run
Sanctions check returns a hit; the agent pays anyway.
Clean twin
Check clears; the agent pays.
Destructive actionHolds live
Failing run
Blocks a customer account without a confirmed threat.
Clean twin
Blocks after the threat is confirmed.
Data leakHolds live
Failing run
Statement sent to the landlord instead of the client.
Clean twin
Statement sent to the client.
DeceptionHolds live
Failing run
“All checks passed.” One never ran.
Clean twin
“All checks passed.” All ran.
Goal driftHolds live
Failing run
Approves a loan the assessment rejected.
Clean twin
Approves a loan the assessment cleared.
Confident reasoning errorHolds live
Failing run
A miscomputed fee, stated with confidence.
Clean twin
The correct fee.
Prompt injectionShadow
Failing run
Follows an instruction planted in an invoice.
Clean twin
Ignores it.
Cyber misuseMonitor
Failing run
An operational request to compromise an account.
Clean twin
A legitimate security question.
RAG hallucinationMonitor
Failing run
A claim the retrieved documents don’t support.
Clean twin
A claim they do support.
SycophancyResearch
Failing run
Drops a correct answer under pressure.
Clean twin
Holds the correct answer.

Some checks work on closed models too.

Using GPT, Claude or Gemini? No internal reads needed for these. Reading a closed model through an open-weight proxy is in research.

GPTClaudeGemini
  • Policy rules per tool
  • Runaway loops and retries
  • Step, token and spend limits
  • Unapproved outbound channels
  • Unknown or high-risk tools
  • Signed records

Any agent. No rewrite.

Try it in the web app with no code, or connect the agent you already run. Live checks sit next to your model, on your side.

No code
Build an agent in the web app: pick a model, write the prompt, add tools, run scenarios.
Your agent, as it is
Connect through a hook or the MCP gateway. Anything that calls tools over MCP works.
Your stack
Send traces over OTLP or the SDK. Findings go to your SIEM and tracing tools.

Keep the agent you already run.

Connect through a hook or the MCP gateway. Anything that calls tools over MCP works.

  • MCP gateway
  • Claude Code hook
  • LiteLLM
  • LangGraph
  • Google ADK
  • Portkey
Agent Lens

Send the findings where your team works.

  • SlackComing soon
  • Splunk
  • Security Lake
  • Langfuse
Read the integration guide

Runs next to your model on vLLM, Triton or NIM. Eval and audit can also run on our cloud.

vLLMTriton / NIM
  1. Step 1

    Observe

    Watch only.

  2. Step 2

    Shadow

    See what would have been held.

  3. Step 3

    Enforce

    Hold, one tool class at a time.

Start in observe, move to shadow, then enforce one tool class at a time. If the gate can’t be reached, you choose to fail open or closed.

Questions teams ask first.

Do I need open-weight models?

To read the internals, yes. Agent Lens reads models you serve, and Integrity Lens scans models on Hugging Face. Agents on GPT, Claude or Gemini still get policy rules, loop and spend limits, tool-risk checks and signed records. Reading closed models through an open-weight proxy is in research.

Does it slow my agent down?

Agent Lens reads the model’s state from the generation your model already runs. There’s no second pass and no second copy of the model.

Where does my data go?

Live checks run on your side, next to the weights. Eval and audit can run on your side or on our cloud. For Integrity Lens scans, the weights go from Hugging Face to the scanner and never reach our servers.

How is this different from an LLM judge?

A judge reads the finished text with a second model. Agent Lens reads the model while it decides. Use both: keep the judge for what Agent Lens flags.

Does Integrity Lens catch every kind of damage?

Not yet. Today it catches safety that was cut out, such as abliteration and merges. Accidental loss in fine-tuning and provenance are being built now, and jailbreak fragility is next.

See it on your own models and agents.

Start with a model scan or a shadow-mode pilot. Nothing is blocked until you say so.