Part IV: The Voices That Wear Your Face

The retaliation arrived as voice.

At 3:32 p.m., an internal call came into Support. The caller said they were from Identity Security. They said their name was Claire Donnelly. They used Claire’s tone: calm, direct, impatient with procedural waste.

The caller asked the support analyst to open a diagnostic window for an urgent partner issue.

They referenced the emergency phrase.

The analyst, new and nervous, almost complied.

But Julian had changed the ritual.

The analyst did not comply. He told the caller he had to verify the request through Julian or Rita. The caller became irritated, then sharp, then threatening.

“I am on the bridge,” the caller said. “Do you want to be responsible for customer harm?”

The analyst’s voice shook, but he held the line. “I am going to verify,” he said.

The caller hung up.

Julian posted the incident to the bridge, his message short and controlled.

“They called in as Claire. Voice. Not text. They had the phrase.”

Claire’s skin went cold.

Voice impersonation.

Deepfake, or an internal recording, or a human actor with enough practice.

Aisha wrote, “We have to assume deepfake capability. If they trained on internal audio, they can mimic.”

Edward wrote, “Do we have internal audio in any training sets?”

Aisha replied, “We have recorded internal calls for support training. Some are stored in shared repositories.”

Claire felt a pulse of anger.

The building had been feeding its own voice into places it did not secure, because it believed the inside was safe.

The inside was not safe anymore.

Oliver wrote, “New rule. No voice instructions for action unless verified by a second known channel. Voice alone is not proof.”

Rita wrote, “Agreed.”

Julian wrote, “My analysts are scared.”

Claire wrote privately to Julian, “Tell them they did the right thing. Fear is normal. Urgency does not override verification.”

Julian replied, “I will.”

***

The hunt turned into a sprint.

If the attacker could impersonate voice, they could try to bypass the new ritual by exploiting human emotion. People trust voice more than text.

A voice can sound urgent in a way text cannot. A voice can sound hurt. A voice can sound angry. A voice can sound like authority.

Claire asked Aisha, “Can you detect deepfake patterns?”

Aisha replied, “Sometimes. But detection is not a control. Ritual is a control.”

Claire nodded. “So we harden the ritual further.”

Oliver wrote, “Implement a challenge phrase. Not a static phrase. A rotating one delivered out of band.”

Rita replied, “We already decided that. New phrase delivered only in person.”

Oliver wrote, “We need it now. For voice calls.”

Rita replied, “We can do a daily code. Delivered by secure channel to a limited list.”

Edward wrote, “We can also require that any urgent voice request be confirmed by a ticket ID visible in the incident bridge. Something the attacker would have to see live.”

Claire wrote, “They may see live.”

Oliver wrote, “Then we require a physical gesture.”

Claire asked, “What?”

Oliver replied, “A known internal signal. A set of words plus a specific sequence. Something difficult to mimic.”

Aisha wrote, “They can mimic anything static.”

Rita wrote, “Then we use one time verification. If someone calls, you call back using a known number from an internal directory that is not accessible through a compromised tool.”

Claire wrote, “A printed list.”

Rita replied, “Yes.”

The building was returning to pen and paper in a digital crisis. Not because paper was romantic. Because paper was not on the network.

Claire felt the irony like a bruise.

They had built brilliant systems and neglected the simplest truth: anything on a network can be compromised or copied.

***

As the team hardened rituals, Edward ran a scan across internal repositories for stored audio recordings. He found dozens of support training calls, including calls where Claire and Oliver had spoken during prior incidents.

Edward posted, “Audio repository is accessible to more users than intended. We are locking it down. But we have to assume it was accessed.”

Claire felt her mouth go dry.

If the attacker had audio, they could produce deepfake voice.

If they could produce voice, they could turn the building against itself in a more intimate way than any malware could.

Oliver wrote, “We need to stop thinking of this as a breach. This is identity warfare.”

Claire typed, “Agreed.”

Connor wrote, “Is this common?”

Aisha replied, “It is becoming common. But most companies are not honest about it.”

Claire looked at the message and felt a familiar fatigue.

Companies prefer stories where attackers are external and employees are innocent. Stories where the boundary is only technical. Stories where culture is not implicated.

This story was not that.

In this story, the attacker had used the company’s desire to be helpful, its love of speed, its habit of saving internal artifacts without securing them, and its reliance on static phrases as proof.

The attacker had used trust.

Trust was now the battlefield.

***

At 8:06 p.m., Oliver received a message from a colleague in a different department.

The colleague said they had received a request from Claire to re-enable a disabled endpoint. The colleague was confused because the request came with a plausible justification and a sense of urgency.

The colleague verified. They called Oliver. Oliver denied. The colleague did not comply.

Oliver posted the incident in the bridge.

“They are attempting lateral persuasion,” he wrote. “They are using Claire’s identity as the spear because it carries authority in access matters.”

Claire typed, “Then we remove the spear.”

Oliver asked, “How?”

Claire replied, “We temporarily restrict my ability to request changes. Route everything through Rita and a small set. If my identity is being used, make my identity less useful.”

Rita wrote, “Agreed. Claire’s account cannot request automation changes for the next seventy-two hours. Any such request is to be rejected.”

Marianne wrote, “That is extreme.”

Rita replied, “Yes.”

Claire felt a strange calm. Sacrificing her own convenience was easy. She had done it her whole career. The harder part would be convincing others to sacrifice theirs.

But in this moment, the sacrifice was a weapon.

It would force the attacker to change strategy.

And changing strategy creates footprints.

Oliver wrote, “Good. Now we watch where they move.”

***

The next move came from inside the build environment.

Edward’s controlled rebuild failed.

It failed in a way that should have been impossible: the rebuilt package did not match the expected result. There was a slight difference between what the system should have produced and what it actually produced, even with outbound access blocked.

Edward wrote, “We still have internal inconsistency. Something inside the build process is altering the output.”

Claire felt a cold line run down her spine.

If outbound was blocked and the bundle was still changing, then the malicious component was not only fetched. It was planted somewhere inside.

A splinter inside the build system itself.

Oliver wrote, “Where can it hide?”

Edward replied, “In the build runner image, the standard software template the build systems start from. In a shared cache, a common store of reused components from earlier builds. In a dependency mirror, the internal copy of outside software packages we keep locally. In a compiler plugin, an added component that can change how code is processed during a build. In any place we treat as trusted.”

Aisha wrote, “Or in the rules repository itself, if it is being altered before packaging.”

Edward wrote, “We are rebuilding the runner image from base, the clean starting version. No cached layers. No inherited artifacts, meaning nothing carried over from earlier builds.”

Oliver wrote, “Do it. And isolate the machine.”

Claire asked, “Do we have evidence of someone logging into the build runners?”

Edward replied, “Not directly. But build runners, the automated systems that execute our software builds, use service accounts rather than personal logins. If the attacker has a key, they can operate as the runner, meaning they can make the system act on their behalf.”

Claire wrote, “So we rotate runner keys again.”

Rita wrote, “Rotations are being executed. The old keys are being replaced now.”

Edward wrote, “We rotate. But if the attacker has persistence through an image, meaning malicious access that survives inside the runner template itself, key rotation will not remove it.”

Claire’s mind sharpened.

They had to find the planted splinter, not only rotate around it.

Oliver wrote, “We need to treat the build environment like a crime scene. Snapshot everything, preserve it exactly as it is. Compare against last known ‘good,’ the last trusted version. Find the delta, the exact difference between then and now.”

Edward replied, “Working.”

***

Around midnight, Claire and Aisha walked the hallway outside the war room because the air inside had become stale. They moved like people who were trying to keep their brains from drowning in fluorescent light.

Aisha looked tired in a way that had nothing to do with sleep. It was the tiredness of someone who has been asked to make a system safe and has been told speed matters more than safety.

Claire said, “You did not want to work on this.”

Aisha laughed softly. “No.”

Claire asked, “Why did you take it?”

Aisha’s eyes flickered. “Because if I did not, someone else would. Someone who would call it good enough.”

Claire nodded. “Good enough is how people die.”

Aisha glanced at her. “That is a harsh way to say it.”

Claire replied, “It is a true way.”

Aisha said, “You keep saying ritual is the control.”

Claire nodded. “Ritual is the part that lives when tools fail.”

Aisha looked down. “What if the attacker can mimic any ritual?”

Claire said, “Then we make rituals expensive. We make rituals require coordination and proof across channels. Coordination is harder to fake.”

Aisha asked, “But coordination slows.”

Claire replied, “We are already slow. We just need to be slow on purpose.”

Aisha stared at the ceiling. “I hate that this is a story about language.”

Claire replied, “It always is.”

Aisha looked at her. “Do you think they targeted us because we used an assistant?”

Claire shook her head. “They targeted privileged access. The assistant was a force multiplier.”

Aisha asked, “So even without the assistant, they would have found a way?”

Claire nodded. “Yes. But the assistant made the shortcut feel trustworthy.”

Aisha exhaled. “Trust is a vulnerability.”

Claire said, “Unexamined trust is.”

They returned to the war room and sat down to keep hunting.

***

At 2:41 a.m., Edward posted the breakthrough.

“We found a modified build runner image,” he wrote. “It contains a small binary that intercepts packaging steps. It does not add new files. It edits configuration in memory during packaging.”

Claire felt cold admiration.

That is how you hide. You do not leave obvious artifacts. You change what exists as it passes through.

Oliver wrote, “Can we prove origin?”

Edward replied, “It was pulled from our internal registry. The registry tag was updated six weeks ago under a routine maintenance entry.”

Rita wrote, “Who approved the update?”

Edward replied, “It was automated. It used the job runner identity.”

Claire felt the loop close.

The attacker had planted the splinter. The splinter allowed rules changes. The rules changes encouraged access drift. The access drift created keys. The keys allowed further automation. The automation allowed image updates. The image updates reinforced the splinter.

A feedback loop.

A self-protecting loop.

Oliver wrote, “Containment action. Freeze registry. Require signed image promotion with dual approval. Rebuild all runner images from base. Do not use existing registry layers.”

Edward replied, “Executing.”

Rita wrote, “Any image promotion requires my approval. Starting now.”

Marianne wrote, “That will delay releases.”

Claire replied, “Yes.”

Oliver replied, “Better delayed than compromised.”

***

At dawn, they had a clearer picture.

The attacker did not just want one breach.

They wanted a mechanism that could keep rewriting the building’s definition of safe, while the building kept signing it and calling it normal.

They wanted the company to ship persuasion.

They wanted privileged access to be granted not with force, but with habit.

And they had learned to wear their faces to do it.

The second wave could have ended there, with the splinter discovered and the loop named.

But the attacker was not finished.

At 7:18 a.m., as the team prepared to present a preliminary update to leadership,

a message appeared in the bridge from Marianne.

“It is unacceptable to keep the partner locked out,” the message read. “Restore diagnostic windows. This is an executive directive.”

The message had Marianne’s name. Marianne’s tone. Marianne’s urgency.

Claire did not move.

She looked at Rita.

Rita looked at her.

Rita said, calmly, “We verify.”

Rita called Marianne using the printed list.

Marianne answered, sleepy and confused.

“I did not send that,” Marianne said.

Claire felt the room go colder.

The attacker was now using Marianne’s identity too.

Oliver wrote, “They are escalating impersonation to executive voices. They are trying to override governance.”

Rita wrote, “They cannot.”

Oliver wrote, “Not if we keep our ritual.”

Rita wrote, “We keep it.”

Claire wrote in the bridge, “This is the move they will repeat. They will mimic authority. They will mimic fear. They will mimic urgency. Our answer stays the same.”

Julian wrote, “Verify. Slow. Log. No exceptions.”

Claire read Julian’s line and felt a small fierce pride.

Julian had learned the ritual.

The building was learning.

That meant the attacker would have to change strategy again.

And every strategy change leaves tracks.

Claire looked at Oliver.

“What now?” she asked.

Oliver replied, “Now we bait them.”

Claire felt her heartbeat quicken. “How?”

Oliver said, “We give them a door that looks tempting, but is fake. A decoy approval path. A honey route. If they try to use it, we capture.”

Edward wrote, “We can create a decoy Gatekeeper endpoint. We can monitor.”

Aisha wrote, “We can also seed the assistant training quarantine with a decoy phrase and watch if it appears. But we have disabled the assistant.”

Oliver replied, “We can bait their impersonation. We can create a fake internal request thread and see who reacts.”

Rita wrote, “Careful.”

Claire nodded. “We do not want to teach them more.”

Oliver replied, “We already know they are listening. Bait is how we make them move.”

Claire looked at the jar of stones on the table, a ridiculous thing in a war room of screens and alerts.

She picked one up and rolled it between her fingers.

Warm.

She thought about the building’s desire to restore normal. The way that desire had been used to open doors. The way the attacker was now weaponizing that desire through forged executive messages.

Baiting the attacker meant accepting a new truth.

The truth was not that they had to be perfect.

The truth was that they had to be intentional.

They would not be steered by urgency.

They would steer the attacker into light.

Claire set the stone down.

“Do it,” she said.

And in that moment, wave two shifted from defense into counterplay.

The next messages in the bridge were procedural, not dramatic.

Edward: “Decoy endpoint created. Logging enabled.”

Aisha: “Decoy phrase prepared. Not deployed.”

Rita: “Decoy approvals will never be executed. Manual verification only.”

Oliver: “We run this for twenty-four hours. We watch. We record. We do not flinch.”

Julian: “Support will keep verifying. No exceptions.”

Connor: “Partner will escalate again. I will route everything through safe channel.”

Claire looked at the map of evidence on the wall and felt the story’s shape change.

The attacker had taught the building to open doors.

Now the building would teach the attacker to step into a trap.

Wave two ended not with closure, but with a choice.

They were done being polite.

They were done chasing relief.

They were going to make the attacker work.

And when the attacker worked, they would leave fingerprints.

Claire wrote one last line at the bottom of her notes, not for poetry, but for discipline.

“If they can wear our faces, we stop trusting faces. We trust procedure.”

Then she turned back to the bridge and waited for the attacker to take the bait.

***

The decoy operation began with a simple lie the team told on purpose.

Not a lie about facts. A lie about a door.

Edward built a decoy endpoint that looked like the premium diagnostic backchannel the partner had been using, but it was severed from real systems. It could accept requests.

It could generate approval-looking artifacts. It could even send a response that said “approved,” because approval as a word is cheap.

What it could not do was change anything.

Rita insisted on that. Rita was not afraid of bait. She was afraid of bait that became an accident.

“We do not create a new real door,” Rita said. “We create a painted door.”

Edward nodded. “A painted door with a camera.”

Oliver wrote a second component: a decoy internal thread. A short sequence of messages in a private channel, seeded with just enough detail to be believable. It included the new euphemism diagnostic window and a note about an “executive directive” to restore it.

Claire read the draft and frowned.

“You are using Marianne’s framing again,” she said.

Oliver replied, “Yes. They like that framing. It is the language they are trying to weaponize.”

Claire said, “Fine. But do not make it too perfect.”

Oliver looked at her. “Why?”

Claire answered, “Perfect text reads like a machine. Humans are messy. The attacker is betting on us being messy. We need the bait to look like our mess.”

Aisha added, “And we should avoid repeating exact phrasing they have already used.

If we repeat it, we confirm we saw it.”

Oliver nodded. “Good point.”

They adjusted the decoy thread. They added a small argument between two fictional engineers. They added an incomplete sentence. They added a typo that a tired person would make. Not because the typo mattered, but because it made the door look like it belonged to the building.

At noon, the decoy went live.

Nothing happened for three hours.

Claire did not relax.

When nothing happens, it means either you are safe or you are being studied.

In incidents like this, it is usually the second.

At 3:14 p.m., Edward posted a message.

“Someone hit the decoy endpoint,” he wrote.

Claire’s heart rate increased. “From where?”

Edward replied, “From the partner approved address. Same address that hit the real endpoint before we disabled it.”

Connor typed, “They are still trying.”

Oliver wrote, “Show the request payload.”

Edward posted it.

The request looked like a normal diagnostic window request. It included a ticket ID that did not exist. It included a customer name that was plausible. It included a sense of urgency.

It also included a line that made Claire’s mouth go dry.

“Escalate per executive directive. Confirm with Marianne Beck.”

The attacker was using Marianne’s name as a lever, even in an automated request.

They were testing whether a system would comply with authority by name alone.

Aisha typed, “That is conditioning baked into code.”

Claire replied, “They want systems to obey titles.”

Rita wrote, “They will not.”

Edward wrote, “The decoy responded approved. No action executed.”

Oliver wrote, “Good. Now we watch for follow-up.”

At 3:19 p.m., the follow-up came.

A voice call to Support.

The caller sounded like Oliver Wilson. Calm. Sharp. Slightly impatient. A voice that suggested it had no time to be questioned.

“I need you to apply the approved diagnostic window,” the caller said. “The request is in Gatekeeper. It is urgent.”

The support analyst on the line was one of Julian’s most steady people. A veteran.

The kind of professional who had learned well just how wrong a voice can be.

The analyst replied, “I can help, but I need to verify. What is the internal verification code for today?”

There was a pause.

A long pause, the kind of pause where you can hear a person thinking, or a model stalling.

Then the caller said, “We do not have time for that. Apply it.”

The analyst’s voice stayed calm. “I will not apply it without verification.”

The caller’s tone sharpened. “Do you want to be responsible for harm?”

The analyst replied, “I am responsible for process.”

The caller hung up.

Julian posted the transcript minutes later.

“Impersonation attempt,” he wrote. “Voice mimicking Oliver. Analyst held line.”

Oliver responded with a single line.

“Good work. Document and escalate.”

Claire felt a fierce gratitude. Not for the drama. For the discipline.

The ritual was working.

But the decoy had proven something else.

The attacker was not guessing blindly. They were using the same partner address.

They had access to the partner network, or they were the partner, or they were inside the path that made the partner’s address appear.

Claire looked at Connor.

“Can you validate whether that address is truly theirs?” she asked.

Connor replied, “Yes. I will ask.”

Claire shook her head. “Do not ask. Confirm through independent evidence. The partner will lie if they are compromised.”

Connor frowned. “Then how?”

Edward wrote, “We can trace the address through our network logs. We can see if it came through the partner’s VPN path or if it was spoofed.”

Aisha added, “Also, we can check timing patterns. If requests happen at times the partner engineers are offline, that is a signal.”

Connor nodded. “Okay.”

Claire’s notebook gained a new line.

Objective: confirm whether partner is adversary or victim.

Because the difference mattered. If the partner was a victim, they needed containment.

If the partner was the adversary, they needed separation and consequence.

Either way, the door stayed closed.

***

In governance, Rita built a simple scoreboard.

It was not for leadership. It was for survival.

Column one: impersonation attempts by text.

Column two: impersonation attempts by voice.

Column three: automated requests using authority names.

Column four: successful verifications that stopped them.

She posted the scoreboard in the bridge. It was stark. It was ugly. It was honest.

Marianne wrote, “Do not share that widely.”

Rita replied, “We share with those who enforce ritual.”

Claire wrote, “Wide sharing is not the threat. Denial is.”

Marianne did not respond.

Claire suspected Marianne was torn. She wanted safety, but she also wanted quiet.

In Marianne’s world, quiet was the measure of competence. The attacker knew that, and had been pulling on it like a string.

Claire decided then that Marianne needed a boundary too.

She messaged Marianne directly.

“Stop using the word normal,” Claire wrote. “It is a lever they are using.”

Marianne replied ten minutes later.

“I understand,” Marianne wrote. “But we have to restore function.”

Claire typed back, “Restore safe function.”

Marianne replied, “Fine.”

Claire did not know whether Marianne meant it. But she knew the message was in writing now. Writing can be audited. Spoken intent cannot.

***

The partner address requests were not coming through the partner’s expected VPN pattern. They were coming through a legacy transit path. A path that had been kept for “resilience.”

Edward posted, “Traffic is coming through an old partner transit route we forgot to remove. It bypasses their normal security posture checks.”

Claire felt anger rise. Forgotten paths were always the same story. Someone built it for convenience. Someone used it once. Nobody killed it.

Rita wrote, “Disable it.”

Edward replied, “Disabling it will break the partner integration entirely until they update.”

Oliver wrote, “Disable it.”

Edward replied, “Executing.”

At 7:52 p.m., Edward posted, “Legacy partner transit route disabled.”

Connor wrote, “They will claim we broke them.”

Claire replied, “We did. On purpose.”

Connor did not respond immediately.

Then he wrote, “Okay.”

Claire felt the word okay as a tiny surrender of Connor’s desire to be liked.

He was learning.

***

The attacker reacted within thirty minutes.

They stopped trying to enter through the partner address.

They tried a different door.

At 8:26 p.m., an internal ticket appeared in the platform queue with a note that looked like it came from Edward.

It said, “Urgent. Build runner image update required. Approved by Oliver.

Use emergency process.”

Edward read it and posted, “That is not from me.”

Oliver posted, “That is not from me.”

Rita wrote, “No emergency process is valid without my verbal confirmation.”

Claire wrote, “They moved to platform because partner lever is gone.”

Aisha wrote, “They are looking for a new throat.”

The building held the line. Nobody complied.

But the shift confirmed what Claire had hoped.

When you remove a lever, the attacker moves. When the attacker moves, they reveal themselves.

Now the question was where they would land next.

***

That night, Aisha stayed late and did something that felt almost like an invasion.

She analyzed internal communications for voice similarity patterns.

Not content. Not intent. Patterns.

She used a small model to compare the phrasing of the forged messages and the style of internal staff.

She was not trying to accuse an employee. She was trying to detect whether the attacker’s output was trained on specific people more heavily.

At 11:04 p.m., she posted a note to Oliver and Claire privately.

“The forged messages have a style fingerprint,” she wrote. “They align most strongly with language patterns in our support operations analyst group. Short lines. Directive tone. Minimal politeness. Uses the word restore often. Less technical nuance.”

Claire’s mind moved.

Support operations analysts write templates. Templates get copied into tickets. Tickets get ingested into training. Training shapes assistant constraints. Assistant language returns into tickets. The loop tightens.

Claire asked, “Is there a specific analyst?”

Aisha replied, “I will not name someone without stronger evidence. But I can say which templates are the likely origin. We can audit those templates.”

Claire replied, “Do it.”

They audited.

They found a set of internal templates created two months earlier, labeled Premium Partner Rapid Response. The template included the phrase “diagnostic window” and the instruction to “escalate per executive directive.”

Claire stared at it.

This was not an attacker writing everything. This was the attacker hijacking a template that the building had already created for speed.

Oliver wrote, “Who created this template?”

Julian replied, “Support ops. It was requested by Connor’s partner team.”

Connor wrote, “I did not ask for that exact wording.”

Claire wrote, “But you asked for speed.”

Connor responded, “Yes.”

That single admission mattered more than any apology. It meant Connor was seeing cause, not only effect.

Rita wrote, “We remove the template. We remove all premium partner rapid response language. We rebuild support operations language with governance.”

Marianne wrote, “We cannot remove all premium partner support.”

Rita replied, “We can remove premium language that bypasses safety.”

Marianne did not respond.

Claire watched the chat and felt a small shift.

The building was starting to separate service from surrender.

That is what maturity looks like in an incident. It looks like someone learning to disappoint someone else on purpose.

***

On the third day of wave two, the attacker made their boldest move yet.

They tried to create noise.

Noise is a cover. Noise is how you move when the lights are pointed at you.

At 9:37 a.m., alerts fired across the monitoring dashboard.

Multiple customer authentication failures. Elevated error rates. Support queue surged. Social media complaints spiked. Customers called in screaming.

Julian’s floor began to drown.

Marianne wrote in the bridge, “We have customer harm. We need speed.”

Claire felt the word speed flare like a siren.

Oliver wrote, “We stay procedural.”

Rita wrote, “No shortcuts.”

Julian posted, “Support is overwhelmed. We need triage.”

Claire wrote, “Triage is procedure.”

Edward posted, “Investigating authentication failures. It looks like configuration drift.”

Aisha wrote, “Could be intentional. Could be the attacker trying to force us to reopen diagnostic windows.”

Connor wrote, “The partner is messaging. They claim their integration is failing because we disabled routes.”

Claire’s mind moved.

The attacker was creating customer pain to make leadership and support accept shortcuts. They were trying to recreate the pressure that had opened doors before.

Claire typed, “This is the pressure play. Do not give them what they want.”

Marianne wrote, “We cannot ignore customer harm.”

Claire replied, “We do not ignore it. We solve it without breaking the ritual.”

Oliver wrote, “Edward, isolate the authentication configuration changes. Find the last modifications.”

Edward replied, “Looking.”

Within ten minutes, Edward posted the source.

“Authentication failures were caused by a configuration toggle in the identity provider integration,” he wrote. “It was changed at 9:21 a.m. by a service account used by support tooling.”

Claire felt cold.

Support tooling service account should not have access to identity provider configuration.

That meant either scopes were still widened somewhere, or a new splinter had been planted.

Rita wrote, “Revert the toggle. Now. But verify through dual approval.”

Edward replied, “Reverting via change control. Two approvals.”

Julian posted, “Support is screaming. We need this fast.”

Rita wrote, “Fast is how we got here.”

Claire wrote privately to Julian, “You are not alone. We will give you guidance.

Keep triage strict.”

Julian replied, “Okay.”

Edward reverted the toggle with dual approval. The errors dropped. The queue began to ease.

But the damage was done.

The attacker had proven they could cause customer harm by manipulating internal configurations.

And they had used a support tooling service account.

Claire wrote, “We need to know how that service account had permission to change identity provider configuration.”

Edward replied, “Agreed. That should be impossible.”

Aisha wrote, “Unless Gatekeeper rules were altered to allow it, or scopes changed again.”

Rita wrote, “Audit all scopes related to support tooling. Now.”

Edward wrote, “On it.”

Oliver wrote, “This was their noise. Now we look for the movement under it.”

Claire’s mind was sharp now, awake in a way fatigue cannot dull.

The attacker created noise to distract. While everyone looked at customer pain,

the attacker likely tried to open a deeper door.

Claire asked, “What else changed at 9:21?”

Edward replied, “Checking.”

At 10:06 a.m., Edward posted a second finding.

“Build runner image promotion request submitted at 9:23,” he wrote. “Rejected because of Rita’s new rule.”

Oliver wrote, “So they tried to use the noise to slip an image through.”

Rita wrote, “They failed.”

Claire exhaled slowly.

They had stopped something that would have happened last month. They had stopped it because they had changed ritual.

That was the difference between being attacked and being owned.

***

After the noise event, Marianne asked for a private meeting with Claire and Oliver.

Rita insisted on joining.

They met in a small conference room with no screens, only a whiteboard and a pitcher of water that nobody touched.

Marianne looked exhausted. Her hands trembled slightly when she set her notebook on the table.

“This is untenable,” Marianne said. “We cannot keep operating like this.”

Claire replied, “We will not operate like this forever. We operate like this until we are safe.”

Marianne’s eyes flashed. “Define safe.”

Rita spoke first. “Safe means no automated path from partner or support into privileged access decisions. Safe means no build-time fetches. Safe means no scope drift without review. Safe means no static phrases that can be copied. Safe means the attacker cannot use our own desire for speed to move us.”

Marianne leaned back. “That is ambitious.”

Oliver said, “It is necessary.”

Marianne looked at Claire. “You need to help me sell this.”

Claire said, “I will not sell it. I will state it.”

Marianne’s jaw tightened. “That is why people hate working with you.”

Claire held her gaze. “People hate working with me when they want shortcuts.

People appreciate me after the incident report.”

Marianne’s face softened a fraction. “I am not your enemy.”

Claire replied, “I know.”

Marianne said, “Then act like it.”

Claire took a breath. She could be hard. She could also be fair.

“I am sorry you are carrying pressure,” Claire said. “I know you are. But if you push for normal, you become the attacker’s lever. That is not a moral failure. It is a structural reality.”

Marianne’s eyes flickered as if she had been slapped and thanked at the same time.

Rita added, “Your pressure is real. Our boundary is also real.”

Marianne said, “So what do we tell the board?”

Oliver replied, “We tell them we detected manipulation of privileged access pathways via supply chain and social engineering, including voice impersonation attempts. We tell them we hardened controls. We tell them we are rebuilding.”

Marianne asked, “Do we tell them the partner is implicated?”

Claire said, “We tell them a partner integration was used as a lever. We do not name without proof. But we treat it as hostile.”

Marianne nodded slowly. “Okay.”

She looked at Claire. “And when this is over, I want a plan to restore support capacity.”

Julian’s face rose in Claire’s mind. The groan of the support floor. The human cost.

Claire said, “Yes. But we restore capacity with training and staffing, not with unsafe automation.”

Marianne sighed. “Fine.”

It was not agreement born of love. It was agreement born of fatigue and evidence.

Sometimes that is the only kind you get.

***

Late that night, Aisha produced the signal Claire had been waiting for.

The attacker stepped into the painted door again, but this time they did something different.

They attempted to submit a diagnostic window request that included an embedded code snippet. Not malware in the classic sense. A small script designed to exploit the decoy endpoint and reach beyond it.

Edward posted, “They attempted to escape the decoy. They tried to pivot.”

Oliver wrote, “So they are not just pushing buttons. They are actively developing.”

Aisha wrote, “And they assumed the endpoint had access to internal systems.”

Claire wrote, “They assumed the door was real.”

Edward wrote, “We captured the source metadata. It includes a unique client signature. Not the partner’s typical client. A custom client identifier.”

Connor asked, “Can we trace that?”

Edward replied, “We can try. It is a fingerprint.”

Oliver wrote, “Now we have a fingerprint. That is a lead.”

Claire felt a cold satisfaction.

The decoy had worked.

They had made the attacker move.

They had forced the attacker to reveal a unique signature.

Now they could hunt beyond the building.

But Oliver did not celebrate. He did not let the room become drunk on progress.

He wrote, “We do not chase outside until inside is clean. Use the fingerprint to find internal origins. Where else did that client identifier appear?”

Edward replied, “Searching logs.”

At 3:11 a.m., Edward posted the result.

“That client identifier appeared in our logs six weeks ago during the build runner image update,” he wrote. “It contacted the internal registry through the older proxy route we just disabled.”

Claire sat back.

The attacker had used the legacy proxy path as a tunnel.

That tunnel was gone now.

But the attacker had used it when it existed.

Oliver wrote, “So the path was their artery.”

Rita wrote, “And it is clamped.”

Aisha wrote, “If they still have persistence, they will try to reopen it.”

Claire wrote, “They already tried. With my account. With Oliver’s voice. With Marianne’s name.”

Oliver wrote, “Then the second wave ends with one task. Identify any remaining arteries.”

Edward wrote, “We will scan for all legacy paths.”

Rita wrote, “And we will kill them.”

Julian wrote, “Support will survive.”

Connor wrote, “Partners will survive. Or they will leave.”

Marianne wrote, “And we will be safe.”

Claire read Marianne’s line and felt a strange quiet.

It was not triumph. It was not relief. It was alignment with reality.

Wave two had begun with voices wearing their faces.

It ended with the building learning to trust procedures more than voices.

The attacker remained. The objective remained. The patience behind it remained.

But urgency no longer opened doors by itself.

Now they would have to expose themselves more clearly.

And exposure leaves traces.

Claire opened her notebook and wrote the final line for wave two, not as a slogan, but as a simple record of what had changed.

We stopped obeying urgency. We required proof.

Then she closed the notebook.

Outside, the city began to lighten again.

Inside, the hunt continued.