Book a demo →
PROTEUS

AEGISAI LAB · MODEL RELEASE

Introducing Proteus

A foundational model for email security that investigates every message before it reaches a person.

Today we are releasing Proteus, the first model from the AegisAI Lab. Proteus is a foundational model trained on email, and it detects attacks by investigating them. It doesn’t rely on rules or behavioral patterns. It checks the claims each message makes against evidence the sender does not control. Proteus does this for every message, every time.

AegisAI began with an observation our founding team made at Google: the diversity of the email attacks we saw every day was accelerating, and detection rules and behavioral models could not keep pace. Over the past year we moved from prompting frontier models, to tuning and then training open-weight models. Each step helped. None had the task specificity needed against attackers who are themselves using generative models. So, we trained our own model, from the ground up.

Watch Ryan Luo and Yogesh Patil on Proteus · 1:55

30x Fewer Missed Attacks, 7x Faster Verdicts

Investigating every attack is what makes the difference in detection. On a benchmark of 3,020 human-reviewed phishing attacks drawn from one week of production traffic, Proteus identified more than 99% of attacks. A leading closed-source general-purpose model identified 81.3%, and a widely used open-weight model 80.0%. Each general model let through roughly thirty times as many attacks as Proteus. Proteus was also faster. Against the closed-source model it returned its verdict 7.2x sooner at the median, and 26x sooner in the slowest 1 in 20 cases.

The general models received the message and a plain classification instruction, which is what a security team gets when it points one at its mail. That makes the comparison realistic, but it is not a measurement of the most those models could be made to do. We set out what this evaluation does and does not establish in the Limitations section.

>99%
of 3,020 held-out attacks identified
~30x
fewer missed attacks than general models
7.2x
faster median verdict than the closed-source model
26x
faster in its slowest 1 in 20 verdicts

Fewer missed attacks, faster verdicts

Proteus against general-purpose models on 3,020 held-out, human-reviewed attacks. Share of attacks missed.

ACCURACY · ATTACKS MISSEDSPEED · SLOWEST 1 IN 20 VERDICTS0%10%20%Proteus<1%Leading closed-source model18.7%Widely used open-weight model20.0%1x10x20xProteus1xClosed-source model26xmedian 7.2x
ACCURACY · ATTACKS MISSED0%10%20%Proteus<1%Leading closed-source model18.7%Widely used open-weight model20.0%SPEED · SLOWEST 1 IN 201x10x20xProteus1xClosed-source model26xmedian 7.2x

Speed is relative to Proteus, measured against the leading closed-source model.

Attacks are now written by machines

For most of the history of email security, novelty was expensive for the attacker. A convincing pretext, tailored to one company and written in flawless prose, took human effort, so a campaign had to be reused across many recipients to pay for itself. That reuse was the defender’s principal asset: an attack could be observed once, characterized, and blocked everywhere else.

Generative models have removed the constraint. Across 2025, AI-generated spear phishing grew from less than 1% to over 20% of the phishing AegisAI observed, and on its current trajectory it approaches the majority of attacks by 2028. It also works. In a controlled study, fully AI-automated spear phishing drew the same 54% click-through rate as messages written by human experts, against 12% for generic phishing. A message tailored to its recipient now costs close to nothing to produce, and each one is a first occurrence.

Phishing is being written by machines: AI spear phishing grew from about 1% of observed phishing in 2025 to over 20%, projected to reach the majority by 2028; AI-automated spear phishing drew a 54% click-through rate.

Why email security needs its own model

The decision to train a model rather than build on top of one came from a class of attack that established methods handle poorly. They do not fail because they are badly implemented. They fail because of what they take as input.

A representative case: a customer’s Google Workspace account was taken over through a malicious OAuth grant and began sending bid invitations to that customer’s suppliers. Every message passed SPF, DKIM and DMARC, because every message was in fact sent by the account it claimed to come from. Authentication was working exactly as specified. AegisAI flagged the first of these messages 89 minutes before the account-takeover alert fired.

Authentication was not the control that failed here. It was the control that confirmed the attacker’s access. The first attack message was quarantined well before any account-level signal existed, because Proteus did not wait for the account to look compromised. It investigated the message itself, and the evidence around it did not support what the message claimed to be.

Three methods, three failure modes

Production email security is assembled from three families of method. Each fails on the case above for a different reason, and in each case the reason is structural. What they share is that each recognizes the present by comparing it with a memory of the past.

Rules and signatures

A rule encodes a property observed in an attack that has already been seen and classified. Inside the space of that property a rule is exact and inexpensive. Outside it a rule has no behavior at all: generalization is not weak, it is zero by construction. That was an acceptable trade while writing a variant of an attack and writing a rule against it were both human-rate activities. Generative models have made the first machine-rate while the second remains human-rate, and the loop no longer closes.

Rulesets also accumulate a cost that is rarely attributed. Each rule narrows the space of delivered mail, and the false positives that result are spread across the ruleset rather than charged to the rule that caused them. A mature ruleset is therefore hard to unwind, because no single rule can be shown to be the expensive one.

Anomaly and behavioral detection

The second family learns a model of normal activity per sender, per recipient and per organization, and flags departures from it. This is where most machine learning in email security is applied today, and it inherits two problems well documented in the intrusion-detection literature.

The first is the base rate. Attacks are rare relative to legitimate mail, and rare events are unforgiving of false positives in a way accuracy figures conceal. If one message in ten thousand is an attack, a detector with 99% detection and a 1% false-positive rate produces an alert stream in which fewer than one alert in a hundred is real: roughly 102 false alarms for every true one. The binding constraint on a deployed detector is almost never its detection rate. It is the false-positive rate required to make that detection rate usable. Axelsson set this out for intrusion detection in 2000, and email presents a less favorable base rate than the systems he described.

The second is what Sommer and Paxson called the closed-world assumption. Machine learning reliably recognizes what resembles its training distribution; anomaly detection asks for the opposite, to identify what resembles nothing seen before. The variance of benign mail compounds the problem. Notifications, marketing, forwarded threads, mailing lists and people who simply write unusually mean that “normal” is not a tight distribution to depart from, and the gap between statistically unusual and malicious has to be closed by something. In practice, that something is an analyst.

In email these problems take a particularly costly shape. The baseline is learned from a sender’s own history, which places the method’s blind spot exactly where losses are largest. A hijacked mailbox writes from inside the baseline that was learned from it. A first contact has no baseline to depart from. Together they describe the typical high-severity enterprise attack, and because additional data is drawn mostly from cases the method already handles, collecting more of it does not close the gap.

What each method can see, by construction: rules and signatures, behavioral baselines and evidence investigation compared across five attack cases.

There is an adversarial consequence as well. A baseline treats observed history as ground truth, so an attacker holding a compromised account can move it deliberately, in steps that are each unremarkable, until the behavior they intend falls inside it.

Supervised classifiers on engineered features

The third family trains a classifier on features an analyst has specified: domain age, link count, keyword presence, authentication results, attachment type. The feature set encodes the analyst’s model of attacks at the moment it was written. When attacks change, it is the features rather than the weights that go stale first, and refitting cannot recover what the features cannot express. Confirmed attack labels are scarce and costly, which limits how quickly the classifier can move. And because campaigns are highly self-similar, a random train-test split places near-duplicates of the same campaign on both sides and reports a number that does not survive contact with production.

Moving from General Purpose to Task Specific

Pointing a frontier model at the mail stream is the reasonable next thing to try, and we tried it before training anything. Three properties work against it.

Instruction-following. General-purpose models are trained to be cooperative with the text placed in front of them, which is precisely the disposition a phishing message is written to exploit.

Email representation. Proteus is trained to interpret an email classification as a structured entity. Our representation brings together the message body, relevant headers, historical context, sanitized URL and attachment context, and threat intelligence. We trained the model to operate on this representation and introduced custom tokenization alongside it. The central choice is not how we divide the text into tokens, but what we teach the model to interpret: a message, its context, and the relationships between them. This both saves on token cost and improves model accuracy, which we discuss in later sections.

Availability. A model anyone can query is a model an attacker can probe offline, repeatedly and undetected, until a message passes. A model that is not public offers no such affordance.

Investigating every message

Three things follow from training our own. A model pretrained on email learns its representation from traffic far larger than any labeled attack set, which is the scarce resource. It can be retrained at the rate the adversary mutates, not the rate at which an analyst can specify features. And, most importantly, it can be trained to investigate rather than to recognize, which is the subject of the next section.

Every method described above asks some version of the same question: does this message resemble something already known? When the attack is new, rewritten or sent from a legitimate account, the honest answer is no, and the message is delivered. Proteus asks a different question.

A baseline asks whether a message is unusual. Proteus asks whether the evidence supports the claim the message makes about itself.

The first question has no answer for a hijacked mailbox, because the message is not unusual. The second question does. It has an answer for every phishing message, because a phishing message has to make claims to work. It claims to come from a particular party, to continue a relationship, to point somewhere legitimate, and to be a step in a process the recipient knows. Each of those claims can be checked against something the attacker cannot author: where the link actually resolves, whether these two parties have ever corresponded, what the route and headers record, and what the sending infrastructure has done elsewhere. None of those checks requires that the attack has been seen before.

Security teams already know that investigation works. It is how an analyst confirms an attack. The limitation has always been scale: a person can investigate a handful of messages a day, so conventional stacks score all mail on resemblance and reserve investigation for the few that get flagged, usually after the fact. Proteus removes that trade-off. Investigation is not an escalation path. It is how every message is assessed.

Investigation on every message

Conventional stacks score all mail and investigate the exceptions. Proteus investigates everything.

TRIAGE — HOW MOST EMAIL SECURITY WORKS

Score everything, investigate a few, after the fact

All mailscored on resemblance
Flagged subseta small fraction
Analyst reviewafter the fact
Verdictsometimes post-delivery

PROTEUS — INVESTIGATION ON EVERY MESSAGE

Every message investigated, inline, before it reaches a person

Every messageno sampling
Claims extractedsender, relationship, link
Evidence resolvedthe sender cannot author
Verdictbefore delivery

Schematic. Illustrates where investigation happens in each approach, not measured volumes or times.

Consider the hardest case in our benchmark, a reconnaissance message. It carries no link and no attachment, it passes authentication, and it asks only for a reply. A content filter has nothing to examine.

Anatomy of an investigation

The reconnaissance message above has no link and no payload. Proteus decides what to check as it goes.

EACH FINDING DECIDES THE NEXT CHECK

active pathconsidered, not pursued

  1. 01Read messageClaims a board member; asks for discretion
    • Link resolutionno links present
    • Attachment analysisno attachments
  2. 02Sender historyNever written from this address
  3. 03Known identitiesBoard member’s known address is not this one
    • Infrastructureauthenticates; adds little
  4. 04Thread historyNew thread, no prior context between them
  5. 05Request patternSecrecy + deadline + reply-first: BEC pretext
    • Payment detailsnone requested yet
  6. VERDICTPhishingBEC reconnaissance

EVIDENCE AGAINST THE MESSAGE’S CLAIMS

bar for a verdict0102030405bar cleared: investigation stops

Illustrative reconstruction of one run. A different message, or a different finding at any step, produces a different path.

Investigated, the message falls apart. A board member who has never written from this address, a new thread with no history behind it, a request for discretion paired with a deadline, and authentication that proves only that someone controls a personal mailbox. No single fact is conclusive. Together they do not support what the message claims to be, and the verdict follows from the evidence rather than from resemblance to anything seen before.

The rest of this post describes how Proteus is built to do this, and what it does to detection when it is done on every message.

Model and training

Evidence ablation

We tested three configurations that differ in one variable: evidence trained into the model, evidence supplied at inference, and no evidence. Supplying evidence at inference identified 8.4% more phishing than serving none. This is the most direct measurement in this post of what investigation contributes: the same model, with and without the ability to check what a message claims. Training it in performed slightly worse than omitting it altogether. We do not have a complete account of that second result. Our working explanation is that training on evidence teaches the model to reconstruct it from the message rather than to consult it, which is to say, to recognize rather than investigate.

Consulting evidence beats memorizing it

Three configurations, one variable, scored on the same pinned message set.

−2%0+2%+4%+6%+8%Evidence supplied at inference+8.4%Evidence trained into the modelslightly below baselineNo evidence (baseline)reference
−2%0+2%+4%+6%+8%Evidence supplied at inference+8.4%Evidence trained into the modelslightly below baselineNo evidence (baseline)reference

Change in phishing identified, relative to serving no evidence.

Calibration

Most models today are tools to aid in decision-making, but don't make the actual decision. These models can produce a confident answer even without necessarily being correct, which makes them impossible to use in a classification context. AegisAI calibrates our models by examining internal activations and measuring how those signals relate to ground-truth outcomes. This gives us an estimate of a decision’s reliability that is separate from how confidently the model states it. The same signals help identify cases where prompt injection has shifted the model away from its assigned task.

82%
of scored messages fall in a usable confidence band
0.037
expected calibration error, before any post-hoc fitting

One practical note for anyone reproducing this: fit calibration on the quantized serving model rather than the training stack. Fitting on the training stack leaves a shift that surfaces later and is easily misread as drift.

Evaluation

The benchmark is a set of 3,020 phishing attacks drawn from one week of production traffic, each confirmed by human review and scored by all three models under identical conditions. None of the 3,020 appeared in training. A miss is any verdict short of phishing, the verdict that quarantines a message and raises an alert, so a message classified as spam counts as a miss.

How the benchmark was built: one week of production traffic, human review, 3,020 attacks with none seen in training, three models, scored.

Accuracy

Proteus identified more than 99% of the held-out attacks. The closed-source baseline identified 81.3% and the open-weight baseline 80.0%. In production, Proteus’s observed precision is 99.996%. This typically equals 1 false positive every 2 years for most companies.

Phishing caught on the benchmark

Share of 3,020 human-reviewed attacks from one week of production traffic.

0%25%50%75%100%Proteus>99%Leading closed-source model81.3%Widely used open-weight model80.0%
0%25%50%75%100%Proteus>99%Leading closed-source model81.3%Widely used open-weight model80.0%

The general models saw exactly the same messages. What they did not do was investigate them. Given the text and an instruction, a general model can only judge whether a message reads like phishing, and a well-written attack does not. Proteus checks what the message claims against evidence resolved at the moment of decision, which is why the attacks that read as legitimate are the ones it catches and the general models miss.

Attacks missed, relative to Proteus

For every attack Proteus misses, how many each general-purpose model misses.

ProteusLeading closed-source modelWidely used open-weight model

PROTEUS = 1x

All attacks
1x
31x
34x
Credential theft
1x
26x
29x
Business email compromise
1x
11x
19x

The gap varies by attack type, and it follows what investigation has to check in each:

  • Credential theft: 26x and 29x more misses. Messages directing the recipient to a collection page that imitates a login form. The deception sits in the link and its destination, which Proteus resolves directly.
  • Business email compromise: 11x and 19x. A claimed executive or supplier requesting a payment or a change of bank details. Here the evidence is relational: whether these parties have corresponded, and whether this request fits that history.
  • Reconnaissance. The first message in a sequence, carrying no link and no payload, written only to establish that the recipient will reply. There is nothing for a content filter to examine, but, as the anatomy above shows, there is plenty to investigate.

Speed

Speed matters more in email than in most applications, because email security sits in the delivery path: a message waits until it has a verdict. Proteus returns that verdict 7.2x faster than the closed-source baseline at the median. The gap is widest in the tail, where the closed-source model’s slowest 1 in 20 verdicts take 26 times as long.

The speed follows from the same specialization as the accuracy. A three-way verdict does not need a model sized for open-ended generation. It is one forward pass with no sampling, over a shorter sequence because the tokenizer fits the input. The tail is a separate matter. A rented frontier model is served to many tenants, with queueing the caller does not control; a self-hosted model has a tail its operator owns. In a blocking path, the tail is what determines whether a detector can sit inline at all, and sitting inline is what allows every message to be investigated before it reaches a person.

Time to a verdict

Relative to Proteus. Bars extend to the slowest 1 in 20 verdicts; the diamond marks the median.

1x5x10x15x20x25xLeading closed-source modelmedian 7.2x26xProteus1x
1x5x10x15x20x25xLeading closed-source modelmedian 7.2x26xProteus1x

In a blocking path the slowest verdicts, not the median, determine how long a message waits before delivery.

Full results

LEADING CLOSED-SOURCEWIDELY USED OPEN-WEIGHTPROTEUS
Attacks identified81.3%80.0%>99%
Share of attacks missed18.7%20.0%<1%
Attacks missed, relative to Proteus31x34x1x
Credential theft missed26x29x1x
Business email compromise missed11x19x1x
Median time to a verdict7.2x—1x
Slowest 1 in 20 verdicts26x—1x

Speed was measured against the leading closed-source model.

Limitations

The results above measure one quantity under one set of conditions. Four things they do not establish are worth stating directly.

The benchmark measures missed attacks, not false alarms. A classifier that labeled every message as phishing would score perfectly on it. False-positive rate is the constraint that determines whether a detector is deployable at all, as the base-rate argument above makes clear. The precision reported above comes from production inference.

The baselines are unoptimized. Each received the message and a plain classification instruction. A team that invested in prompt design, retrieval of surrounding evidence, or fine-tuning on labeled attacks would close some part of this gap, and we do not know how much. The comparison describes what a general-purpose model does out of the box, which is the decision most security teams are actually weighing. It is not a claim about the limit of what such a model could be made to do.

The sample is one week of traffic from our own customers. The attacks reaching a different set of organizations will differ, and a single week does not capture seasonal or campaign-level variation. We would not expect the exact ratios to reproduce on another sample.

Detection performance is not stationary. These figures describe a model measured against traffic that had not yet adapted to it. The release process below shortens the interval over which decay accrues. No release process removes it.

How we ship Proteus

Because detection decays, how a model is updated matters as much as how it scores at launch. Every Proteus version is continually retrained. Each version must clear an adversarial suite covering nine attack families, and then runs beside the version it would replace on live traffic. It is promoted only if it does better there.

How a Proteus version reaches production: observe, retrain, adversarial suite, shadow beside its predecessor, promote only if it wins.

The defender’s window

This reading of the next few years is not ours alone. The laboratories building frontier models have arrived at a similar account from the opposite direction.

OpenAI has described a “defender’s window”: a period in which the capabilities that let attackers find and exploit flaws at scale are equally available to defenders, and which narrows as the offensive side industrializes first. Greg Brockman’s framing is that the window “is open now,” and that over the coming months every organization will need to begin significantly automating its security program. Anthropic has approached the same problem structurally, applying cybersecurity safeguards to its strongest models while operating a Cyber Verification Program that grants vetted defenders tiered access to capabilities that are otherwise restricted.

Both positions rest on an assumption worth making explicit: access to a capable general model is a precondition for modern defense, not a defense in itself. Proposals to slow AI development are unlikely to succeed; models are globally distributed technology, and attackers have already adopted them. The remaining path is to build defenses that improve at the same rate. What frontier capability is spent on is a separate decision, and general capability applied generically to a specialized adversarial problem is measurably weaker than capability spent on a model built for it. Our benchmark is one measurement of that difference, in one domain.

Email is where the difference matters most. It remains the principal entry point for enterprise compromise, and it is the one surface every employee is required to keep open.

What comes next

Proteus is the first model released by the AegisAI Lab. The direction is a family of models and the investigation tooling around them, so that every message reaching a person is assessed on evidence rather than scored on its resemblance to past attacks.

About the Authors

  • Bin Lu is a Principal AI Engineer at AegisAI. He holds a Ph.D. from CUHK/Cornell and formerly served as Engineering Lead for Chrome Enterprise Security.
  • Pavel Petrenko is the Threat Intelligence Lead at AegisAI, formerly serving as Google Ads Security Lead.
  • Sam Malik leads Adversarial AI at AegisAI and previously led GCP AI Platform Security.
  • Yogesh Patil is a Founding AI Engineer at AegisAI and holds an M.Eng. from Cornell University.
  • Ryan Luo is Chief Technology Officer at AegisAI and formerly served as Google reCAPTCHA/Safe Browsing Lead.

References

  • Axelsson, S. (2000). The base-rate fallacy and the difficulty of intrusion detection. ACM Transactions on Information and System Security, 3(3), 186–205.
  • Heiding, F., et al. (2024). Evaluating large language models’ capability to launch fully automated spear phishing campaigns.
  • Sommer, R., and Paxson, V. (2010). Outside the Closed World: On Using Machine Learning for Network Intrusion Detection. IEEE Symposium on Security and Privacy.
  • OpenAI. The Defender’s Window. openai.com/index/the-defenders-window
  • OpenAI. Cybersecurity in the Intelligence Age. openai.com/index/cybersecurity-in-the-intelligence-age
← All posts

Introducing Proteus

A foundational model for email security that investigates every message before it reaches a person.

Introducing Proteus
Newsletter

Don't miss the next big threat

New attack write-ups, product changes and practical defence notes, sent when there is something worth reading.

Keep reading

Related posts

More from the AegisAI blog.

No related posts yet.